X Archive
-
OpenAI is following the ASPAVA strategy, with gamified subsidies ASPAVA is a type of kebab shop from turkey. These shops usually serve mid quality food But despite that, the majority loves them. That's because they serve free stuff throughout They serve free sides at the beginning. They serve free sides during the main course. They serve free dessert at the end. They serve free tea as many times as you like Some of them give you so much "free" food that you feel like you are stealing. A lot of people frequent these shops not because they like it a lot, but because they are addicted to getting things for free Main dish portions are made smaller in order to breakeven with so many side dishes, which I feel is analogous to what is happening with GPT these days---though I can't prove it NVIDIA is not a car, and OpenAI is not kebab. I don't know if codex resets cost OpenAI much, but if they are, then they might be a waste of resources Developers are the most disloyal customer group. Once the subsidies are gone, they can switch away in the blink of an eye -
-
-
-
Most people, including professionals, will likely NOT pay more than $10k capex for their home AI workstation, even in the long run (ignoring future inflation) If you stretch it, <= $20k They will also not want to pay more than $200~300/mo opex, on electricity for example Builders who are spending a lot more than that on hardware: keep that in mind, if you are building for the general public dogfood your product with that which everyone will have, not just 0.1%
-
Is anyone able to run nvidia/Qwen3.6-35B-A3B-NVFP4 with the config suggested in the readme? It OOMs before it can start serving huggingface.co/nvidia/Qwen3.6…Image hidden
-
Is anyone able to run nvidia/Qwen3.6-35B-A3B-NVFP4 with the config suggested in the readme? It OOMs before it can start serving huggingface.co/nvidia/Qwen3.6…Image hidden
-
-
I know that my macbook local model benchmarks have started when my lap catches on fire Add to this list: "does not set the room on fire"
-
-
I quite like this token speed simulator by @mikeveerman And I keep losing it, so hopefully I will remember to come back to this tweet:) Link: mikeveerman.github.io/tokenspeed
-
I keep seeing insanely expensive builds giving insanely impressive results These results don't matter What matters is, whether one can: - run a "SOTA level" model, whatever that is - with under 32gb VRAM or unified memory - in 5 parallel sessions - with 50~100 tok/s each - with enough leeway memory for other applications - in a system as cheap as $1000 That is our goalpost That is the threshold when open source AI will win
-
-
Another annoying GPT-ism (circa June 2026): while describing something, it always describes what it *does*, but never what it *is* > "LocalPerf benchmarks local LLM inference servers and keeps the evidence in one portable run artifact." If I were a philosopher, I would say that "AI lacks ontology". Or that it is "anti-essentialist", believes that things cannot be things in themselves But all models literally have an ontology. They have it since word2vec days, you can plot it out. It's just an annoying tendency in GPT's writing So using philosophy to understand AI might be dumb sometimes If you don't want your README's to sound like slop, then you can steal my write-readme skill: After write-readme: "LocalPerf is a local LLM inference benchmark CLI. It runs benchmark plans against local inference servers and stores the evidence in one portable run artifact." Skill: github.com/osolmaz/tools/blob/main/agents/skill…Image hidden
-
Give @LottoLabs a follow if you are not already He is building localmaxxing.com, crowdsourced LLM benchmark results and performance profilings Very very coolImage hidden
-
This. Especially when the whole machine hangs instead of OOMing when my agent accidentally loads too many models into memory (lmk if there is a firmware update or sth that fixes this on the spark now, creating a cgroup doesn’t work) forums.developer.nvidia.com/t/dgx-spark-be…
-
Apple hiked prices as I was posting this 💀
-
Was great to be there, thanks for the invite @lionelsimai!
-
How I imagine @TheAhmadOsman I actually bought my GB10 thanks to him back in feb, at a discount Give him a follow if you are not already!Image hidden
-
The Local Frontier is advancing The amount of AI memory for inference we can get for less than $3000 has been steadily increasing The memory crunch has slowed this down and even made it retrograde. However, once we bounce back from it, the progress will be gloriousImage hidden
-
Was great to talk, thank you @ben_burtenshaw for inviting me!
-
I knew it would find me, sooner or later 🫠
-
Besides being hot, this take is very correct and points out to a fundamental tradeoff in the storage layer being centralized versus distributed git is distributed and that makes total sense for code which takes small space by its nature. it is cheap for everyone to duplicate it locally. this proved to be very useful over e.g svn, when devs could develop independently from the centralized server AI artifacts, however, are 1 million times bigger than code. in that case, the bottleneck becomes storage and network. decentralization and version control become lower priority. they can be sacrificed the tradeoff tilts towards getting the cheapest possible storage and transfer. because you will need a LOT of that I regret to announce to competitors that Hugging Face has already won this game when they acquired Xet. the tech just works, and the network effects are immense
-
Awesome. A feat unimaginable before agents
-
My posts here on X now sync automatically to my blog, giving me full ownership of my content and zero effort SEO For free, no API costs My long-form posts are automatically featured and titled on the front page of solmaz [dot] io. Filtering is done by my claw running a sync-x skill daily, which then notifies me on Discord How do I scrape the posts? ALL the posts I view (including the ones I post) are saved locally using @kubmi's xTap and then synced to a private repo using my extension on it xtap-sync My claw has access to that private repo and can run programmatic tasks like sync-x, summarize what happened that day, notify me about any topics I wantImage hiddenImage hiddenImage hidden
-
My posts here on X now sync automatically to my blog, giving me full ownership of my content and zero effort SEO For free, no API costs My long-form posts are automatically featured and titled on the front page of solmaz [dot] io. Filtering is done by my claw running a sync-x skill daily, which then notifies me on Discord How do I scrape the posts? ALL the posts I view (including the ones I post) are saved locally using @kubmi's xTap and then synced to a private repo using my extension on it xtap-sync My claw has access to that private repo and can run programmatic tasks like sync-x, summarize what happened that day, notify me about any topics I wantImage hiddenImage hiddenImage hidden
-
-
Excited!
-
If you are interested in running such demos, look into --demo mode in my local model swiss army knife localpi github.com/dutifuldev/loc… Thank you @googlegemma for the shoutout
-
Huge. Added my github TUI github.com/dutifuldev/ghz…Image hidden
-
Huge. Added my github TUI github.com/dutifuldev/ghz…Image hidden
-
-
Will talk about my recent adventures running local models on the Spark, Pi and OpenClaw Click Notify Me on youtube to stay tuned
-
New blog post: Using local models for agentic zero-shot classification, in real-time, high frequency triage If you have a 128gb of memory for models (a DGX spark like I do for example), you can create a real time classifier and notifier for yourself that can classify more than >20 items per minute, using mid-sized @googlegemma and @Alibaba_Qwen models, with over 200-300 output tok/s aggregate throughput Like processing new tweets on twitter, issues/prs on github, messages on telegram and discord, in real-time Over the past few weeks, I have built one for myself, to filter and get notified about local model related issues on the OpenClaw repo I initially thought gemma-4-e4b would give me the best tradeoff I was wrong. I learned that if one has enough memory already, one should not bother with <10b models like gemma4 e4b or e2b. Precision and recall were much higher zero-shot with gemma-4-26b-a4b, whereas the smaller e4b needed significant prompt optimization to eventually not perform nearly as good To provide more context to the model, I created a restricted bash-like shell, called reposhell. In that shell, it can run read-only commands to ls/find/grep/cat openclaw source code, but only that. When the PR description/diffs are not clear enough as to categorize it, the agent reads the code to figure it out Because small models can get prompt injected, and I need to make sure that someone can't harm my setup by creating a malicious issue or PR in the openclaw repo I found that for specific systems like this, it is very convenient to extend and bundle Pi. You can create agentic CLI tools that work fully locally and for free, and keep that separate from your main pi coding setup. localpager-agent has its own session dir and tools, and I ensure that it will run local models in a secure way by isolating it from my main pi setup Once localpager-agent categorizes a PR/issue as local_models and related labels, I automatically receive it as a notification on Discord The whole implementation is fully open source and MIT licensed, alongside the dataset we used to benchmark the performance I believe zero-shot agentic classification running on local hardware will find many use cases across a wide variety of business applications, like news gathering, open source software development, customer support, content moderation, sales and so on Agents increase the amount of information produced in a lot of systems, and hence we will need to set up cheap ways to wrangle all that information In times where governments can cut off access to SOTA models on a whim, it is more important than ever to build your business on open models and if possible, run them on your own hardware! Big thanks to @evalstate and @ben_burtenshaw for their valuable feedback, especially with helping me evaluate this more rigorously! One take-away is that categorizing contributions in an open source repo is a *hard* problem, and that it is not trivial to reliably create a golden dataset with LLMs, for evaluation purposes Read more here: huggingface.co/blog/local-models-pr-triage
-
gpt 5.5 is not naturally good at modeling and cannot create simplified nice mathematical models completely autonomously I did a parameter sweep with gemma-4-31b-a4b on memory usage, output tok/s etc. while varying context window, concurrency and other parameters. It took quite a few tries, and I still do not trust the model that gpt5 fit to the data besides, it measured linux cgroup memory and not the actual gpu memory used, so the whole sweep is wasted... output tok/s looks more accurate though, soon I will have a model that can give the optimal parameters over the space of context window <> concurrency <> tok/s <> memory usage off to do another run
-
gpt 5.5 is not naturally good at modeling and cannot create simplified nice mathematical models completely autonomously I did a parameter sweep with gemma-4-31b-a4b on memory usage, output tok/s etc. while varying context window, concurrency and other parameters. It took quite a few tries, and I still do not trust the model that gpt5 fit to the data besides, it measured linux cgroup memory and not the actual gpu memory used, so the whole sweep is wasted... output tok/s looks more accurate though, soon I will have a model that can give the optimal parameters over the space of context window <> concurrency <> tok/s <> memory usage off to do another run
-
-
gpt5.5 and most other models are very bad at one-shotting nice data models gpt5.5 also has this annoying property that once it decides for a schema (or any design), it's very hard to trigger thinking again. and if you ask to "rewrite from scratch", it will write create something even more ridiculous To solve this problem, I have built a meta-harness over codex just for simplifying slop data models called schemator (work in progress) Basic idea: it mimics what I myself do while I am designing a schema: scrutinize and question each field one by one It starts a fresh codex session for each field with a fixed prompt like "Try to come up with the most Lindy data model" + a prompt for side notes It does that with a fresh context for each field, so that they are independent from each other. At the end of a review run over a field, the reviewer can propose to keep, rename or remove the field When all fields are reviewed once, that makes one iteration. Then this is looped over until the review results stabilize, and do not propose any further changes I get better results by just asking my agent to "use schemator on this" after it creates a JSON schema or SQL table Give it a try if you have codex! It has a skill, so should be easy for an agent to figure out how to use github.com/dutifuldev/schematorImage hiddenImage hidden
-
For my recent LLM leaderboard osolmaz-leaderboard.hf.space, I sum up all time total downloads (or likes) across model variants, and then divide it by the age of that model. I.e. "time decay" for popularity This gives a more time-agnostic metric for the popularity of that model. In an ideal ranking, older models that are not popular anymore should be demoted, like 2 year old Llama 3 models. If you don't do that, they might still occupy top 10 needlessly, despite having been replaced by e.g. qwen in practice Thanks to that, qwen-3-6b which came up 1 year ago and has 150m downloads can surpass llama-3-1-8b which came up 2 years ago and has 200m downloads More notes on my post: solmaz.io/popularity-rankingImage hidden
-
if you take the Most Downloaded Models of All Time, Llama 3.1 makes it to Top 10 with around 200 million total downloads (ranking is done w.r. to time-averaged downloads) RIP Llama, you walked so @googlegemma and @Alibaba_Qwen can run Also a reminder that if you build your branding on top of open weight models developed by big corps, you might eventually be the de facto owner of that brand if they pull the plug on it. Like llama.cpp @ggml_org Huge fumble by MetaImage hidden
-
My LLM leaderboard osolmaz-leaderboard.hf.space auto discovers different variants of model releases, even if they are not linked by base_model From this, I found out that @RedHat_AI was the first to release NVFP4 quantization for qwen3-6-35b-a3b Nice to see everything in one placeImage hiddenImage hidden
-
"big token will bless me with free tokens today inshallah" is not a healthy mindset to nurture don't get me wrong, I love the subsidies and the memes@thsottiaux·Image hidden
-
If you are in AI, just don’t be anon here I see a bunch of anon accounts posting great local model content… what’s the point of being anon? To seem cool? Most of those accounts are not doing anything illegal, so there is no point. It would add so much more legitimacy to your work if you just put your real face and not a slop or anime girl pfp it only makes sense for those who are abliterating models. otherwise, it makes you seem sus just put your real face anon
-
I created an LLM leaderboard based on Hugging Face download and like counts, grouped, filtered and time-averaged. Top 5 downloads is shared by @Alibaba_Qwen and @googlegemma 👑🤝👑 Top 5 likes, on the other hand also includes @deepseek_ai V4 Pro 👑 Even @OpenAI makes it to #8 top downloads with gpt-oss-20b 👑 qwen3-6-35b-a3b is the second most CIRCULATED LLM of this year, with an average of 21 million downloads per month, since the day it was released 2 months ago 📈📈📈 Despite first place belonging to 8mo old qwen3-vl-2b-instruct, the highlight belongs to the mid-sized MoE model, which has hit a size/performance sweet spot so hard that it absolutely 💥 SHATTERED 💥 Hugging Face leaderboards in the 2 months since it has launched qwen3-6-35b-a3b is followed closely by its dense sibling 27b --- and then the mid-sized gemma 4 models 26b-a4b and 31b Note that a model's distribution is inversely proportional to its size, but not strictly! Usefulness plays a factor as well, since gemma 4 26b-a4b is being downloaded more than the smaller gemma 4 e4b I created this leaderboard because Hugging Face's all time highest downloads and likes did not give me enough information about what is really popular, neither today, nor all-time. I wanted something in between How do I calculate this ranking? - Get models that with n_downloads >= 100k - Exclude models older than 1 year - Deduplicate and group quantizations and variants of the same model based on slug prefix heuristics - For each group, sum up total downloads of all time - Sort by descending total_downloads / age = average_downloads_per_day (can also sort w.r. to likes per month) - Repeat every day to get the most up to date ranking More info and source on the leaderboard page, hosted on a Hugging Face space: osolmaz-leaderboard.hf.space This is a work in progress, please reply below if you see a model that should be there is missing, or any other mistakesImage hidden
-
-
gemma-4-26b-a4b is the most CIRCULATED LLM of recent history, with an average of 126k downloads per day, since the day it was released 3 months ago Top 10 is shared by Qwen and Gemma, with DeepSeek V4 Pro coming in close 🤝 Note that a model's distribution is inversely proportional to its size, but not strictly! Usefulness plays a factor as well, since gemma 4 26b-a4b is being downloaded more than the smaller gemma 4 e4b I created this leaderboard because Hugging Face's all time highest downloads and likes did not give me enough information about what is really popular *these last few months* How do I calculate this ranking? - Get models that with n_downloads >= 100k - Exclude models older than 1 year - Sort by descending total_downloads / age = average_downloads_per_day (can also sort w.r. to likes per month) - Deduplicate quantizations etc. of the same model based on slug prefix heuristics More info and source on the leaderboard page, hosted on a Hugging Face space: osolmaz-leaderboard.hf.spaceImage hidden
-
I need better UI/UX on queueing messages to agents. I want to be able to: switch the order of queued messages pause the queue edit any message that are still in the queue undo steer messages in the few seconds they are being sent I want more visual emphasis on the queue, like a Queue View I can toggle, that puts the queue at the center I want this in all the UIs and coding agents, codex CLI, desktop, moshi... especially while on the phone
-
saving the world from AI cartelization for fun and profit
-
herdr is my terminal now, both local and remote. highly recommend
-
16x parallel Gemma-4-26B-A4B-NVFP4 runs 🤯🤯🤯 18 output tokens/s, aggregate 300 tok/s 1 DGX Spark with 128 GB unified memory Concurrency so high I had to demo it programmatically It can go up to 32 even! 🤯 But then my screen would not have been readable for you And this is not even using flashinfer yet! Please reply if you know whether support is on the way Note that this is not dumb e4b or e2b that you can run on the average laptop. This is the big Gemma MoE Model link: huggingface.co/nvidia/Gemma-4-26B-A4B-NVFP4
-
-
I did some math, and running my Nvidia GB10 workstation (Asus GX10) costs me maximum: 12~13 USD / month or 150~160 USD / year It is a little bit above half the price of ChatGPT plus subscription. For that, I get to run models that can fit in 128 GB of memory How I calculated: You can see how much power your apartment uses in Singapore in half-hourly resolution. We turned off all devices and A/C while we sleep, and got only the fridge and the GB10 remaining From that, we see it uses around 80-100 Watt while I was running an inference workload overnight. So this is like an upper bound I take it as 90 Watt. Electricity here costs 0.25 SGD / kWh 0.09 * 0.25 * 24 * 30 * (SGD/USD conversion rate) = 12~13 USD / month = 150~160 USD / year Local models are getting very good now, small ones roughly around GPT 5.x-mini level. This workstation makes all sorts of workloads possible for me that would otherwise cost a ton on the API It is also my always on workstation that works overnight. I use Codex for my work, and my workstation is always running agents. It never sleeps. I never have to worry about keeping my laptop lid open. I connect and monitor the agents anytime on my phone using mosh and herdr We have crossed a threshold. Running local models is cheaper than a big token sub for quite a few workloads already. If you are running a business, that makes a difference The localening is here
-
THE LOCALENING IS HERE
-
New agent benchmark alert: SkillsBench
-
nvidia/Qwen3.6-35B-A3B-NVFP4 running in vLLM nightly on my Nvidia GB10 is actually insane 50 tok/s, 4 concurrent generations. total 200 tok/s. ideal for spawning subagents or working in parallel its tool calling behavior is very good as well. I will be giving it test drive on an openclaw instance, and keep you posted More details on NVIDIA forum: forums.developer.nvidia.com/t/benchmark-report-…
-
nvidia/Qwen3.6-35B-A3B-NVFP4 running in vLLM nightly on my Nvidia GB10 is actually insane 50 tok/s, 4 concurrent generations. total 200 tok/s. ideal for spawning subagents or working in parallel its tool calling behavior is very good as well. I will be giving it test drive on an openclaw instance, and keep you posted More details on NVIDIA forum: forums.developer.nvidia.com/t/benchmark-report-…
-
-
Click open GitHub PRs and issues directly in the side pane in @herdrdev, instead of having to go to the browser. As many issues and PRs as you want, WITH TABS! Install ghzinga herdr plugin and just ctrl+click the link: github.com/dutifuldev/ghzinga Thanks @lumendriada for sneaking in the ability to capture link clicks 2 days after I requested it! God I love open source...
-
Hugging Face buckets are very literally, actually, 100%, a game changer note that I never ever use that phrase
-
Trying to copy wrapped URLs is a pain not only in ghostty/iterm2 but also in mobile apps like Moshi On the laptop it’s fine because I can select rectangular area and edit it, or make the window bigger On the phone, its’s impossible. Fingers too big, too much of a hassle Should a terminal emulator try to detect these? It already detects herdr. What do you think @odd_joelImage hidden
-
Current average generation speeds for local DeepSeek-V4-Flash-Q2, highest to lowest: Mac Studio M3 Ultra: 32 tok/s MacBook Pro M5 Max: 30 tok/s Apple ??? M4 Max: 25 tok/s MacBook Pro M3 Max: 24 tok/s Mac Studio M2 Ultra: 22 tok/s NVIDIA DGX Spark / GB10: 13 tok/s It seems macs' higher memory bandwidth is contributing here, though I'm not sure if GB10 performance could be improved (I do hope so, I have one!)
-
We have local Deep Research Now we just need to index the whole internet to have local ChatGPT 😅
-
Btw, TTS has come such a long way, @GoogleDeepMind cooked with gemini-3.1-flash-tts I gave Codex my google credentials and it oneshotted the Gemini TTS implementation When I built this 4 years ago, Azure TTS used to be SOTA. Then @ElevenLabs came in and raised the bar super high. Now Google is going after their lunch with controllable expressiveness at scale. I cheer for both! Here is Manim Voiceover demo from 4 years ago with Gemini TTS (sound on)
-
-
I major concern I have these days is, while I author code in languages I cannot manually code, are they any good? Over years, I have worked with a number of languages: C, C++, Fortran, MATLAB, JavaScript But Python was my go-to language since more than 10 years. Well that changed last summer So while I have strong opinions on how Python code, should be, conventions and all, I don't have so strong opinions on other languages. That means I am producing slop by default in Rust, Go and TypeScript To solve that problem, I created github.com/dutifuldev/slophammer Its aim is to be "the only tool and resource your agent needs, to minimize slop" It is inspired by the recent bathrobe rants of @unclebobmartin, a.k.a. the author of clean code It enforces a minimum test coverage, maximum cyclomatic complexity, mutation tests, code style across different languages But I have a major issue: How do I know that Slophammer itself isn't slop? One way is to implement and use it for Python, the language I know better, and judge what kind of changes it enforces So for this weekend experiment, I used Slophammer to refactor, improve coverage and merge new features to one of my old Python projects, Manim Voiceover github.com/ManimCommunity/manim-voiceover The result is... mixed. We now have types everywhere, which is great. But the constraints have also made it write garbage code like this one. It works fine, even though it's not elegant. The new feature also works What do you think? Does code still need to be aesthetically pleasing to the human eye? Should it still be human readable? If an agent writes slop in the forest, and there is no-one to read it, is it still slop? If anything, I should use its output in Python to reason about other languages, and add more and more constraints. The more the constraints, the less the slopImage hidden
-
This is why I love this site, open collaboration!
-
-
This is what I have been feeling recently as well, looking at models write code better and faster than me
-
Dabbling in GEPA. Codex's /goal on GPT 5.5 high is still surprisingly reward-hacking I had set a /goal before I slept to implement a plan. It ended the loop after doing just 1 iteration It feels like the model is following the path of least resistance and slacking off. Though it could also be me putting "try to make good progress in 8 hour's time" in the prompt, can't be sure Lesson: When you are doing such a solver loop, always specify min_iter and max_iterImage hiddenImage hidden
-
-
Experimenting with SOUL.md on gemma4-26b-a4b (running on @DeepInfra) Interesting that such a lightweight model can already run such a conversation in openclaw harness @GoogleDeepMind cooked hereImage hidden
-
Experimenting with SOUL.md on gemma4-26b-a4b (running on @DeepInfra) Interesting that such a lightweight model can already run such a conversation in openclaw harness @GoogleDeepMind cooked hereImage hidden
-
Experimenting with SOUL.md on gemma4-26b-a4b (running on @DeepInfra) Interesting that such a lightweight model can already run such a conversation in openclaw harness @GoogleDeepMind cooked hereImage hidden
-
Experimenting with SOUL.md on gemma4-26b-a4b (running on @DeepInfra) Interesting that such a lightweight model can already run such a conversation in openclaw harness @GoogleDeepMind cooked hereImage hidden
-
Experimenting with SOUL.md on gemma4-26b-a4b (running on @DeepInfra) Interesting that such a lightweight model can already run such a conversation in openclaw harness @GoogleDeepMind cooked hereImage hidden
-
-
-
RIP 🙏 x.com/yushen686/stat…
-
RIP 🙏 x.com/yushen686/stat…
-
RIP 🙏 x.com/yushen686/stat…
-
RIP 🙏 x.com/yushen686/stat…
-
-
-
-
What did @karpathy see / was shown? Why did the benefactor and teacher of the whole ML ecosystem join Anthropic, a company the polar opposite of his image, on the eve of such a powerful model release It can't be purely money Did he reckon that the only way to benefit humanity was to be on the inside, or rather, to not be left outside, of whatever is brewing in there?
-
-
Fable burned through my 5 hour quota, and then automatically fell back to usage credits without asking. Org settings I suppose It was burning through 1 usd every few seconds It burned through 66 usd before I reacted. Yeah, this is not affordable for anyone with that API pricing, without subsidy/planImage hidden
-
Fable burned through my 5 hour quota, and then automatically fell back to usage credits without asking. Org settings I suppose It was burning through 1 usd every few seconds It burned through 66 usd before I reacted. Yeah, this is not affordable for anyone with that API pricing, without subsidy/planImage hidden
-
-
It’a been a little bit over 1 year since Anthropic released their Max plans and Claude Sonnet and Opus 4, thus making Claude Code affordable and kickstarting the agentic revolution Opus 4 was a glimpse into the future. I’ve spent the entire summer swearing at it and typing ultrathink Today, Fable 5 feels like another step change I no longer need to type ultrathink. And no longer need to swear at Anthropic models. Only at their marketing team.
-
To YOLO with Fable 5, or not to YOLO, that is the question... The last time I left, Claude models still had tendencies to rm -rf your home folder or delete stuff without asking first. Is this still a risk? And from the looks of it, Claude Code still doesn't have Codex's LLM-filtered approval gate feature. Or am I missing something? Please enlighten your fellow Claude noob 😇Image hidden
-
To YOLO with Fable 5, or not to YOLO, that is the question... The last time I left, Claude models still had tendencies to rm -rf your home folder or delete stuff without asking first. Is this still a risk? And from the looks of it, Claude Code still doesn't have Codex's LLM-filtered approval gate feature. Or am I missing something? Please enlighten your fellow Claude noob 😇Image hidden
-
To YOLO with Fable 5, or not to YOLO, that is the question... The last time I left, Claude models still had tendencies to rm -rf your home folder or delete stuff without asking first. Is this still a risk? And from the looks of it, Claude Code still doesn't have Codex's LLM-filtered approval gate feature. Or am I missing something? Please enlighten your fellow Claude noob 😇Image hidden
-
-
Speaking of loops, I have renamed my implementation-loop skill from earlier this year to autoimplement, because it's shorter Calling skills that loop auto-x, auto-y makes them more memorable than calling them x-loop, y-loop But it also increases the number of keystrokes you have to type, before you can tab-complete them Alas, I like still this more github.com/osolmaz/tools/tree/main/agents/skill…Image hidden
-
-
The masculine urge to create your own agent multiplexer
-
Just in time for a lot of Codex-default developers going back to Claude Code momentarily to try out Fable 5 Here is a CLAUDE.md -> AGENTS.md symlinker that should save you from the hurdles of obstinate Anthropic conventions It installs a hook that creates the CLAUDE.md symlink automatically as Claude Code traverses directories that contain AGENTS.md, automatically ignored by git No need to create CLAUDE.md with reference to AGENTS.md like Anthropic suggests. It just works github.com/dutifuldev/claude-md-symlinkerImage hidden
-
-
Question to my ghostty-savvy friends I am trying to reproduce the Quake style dropdown experience I have been using since 2010 on ghostty on mac here. nothing works quite as well as iterm2 yet I tried ghostty quick terminal mode. good but it doesn't let me open multiple tabs I tried cmux because it ships ghostty anyway and is supposed to have more features. but its system-wide hotkey is not playing well with aerospace and window focus iterm2 worked perfectly. tap control double and I'm in the terminal. is there anything that replicates this UX
-
-
ghzinga can now show multiple PRs/issues in tabs natively, no need to create a new pane in tmux/herdr also, you can tell your agent to open all the relevant issues/PRs in a side pane using it, and it should work seamlessly it's the open source maintainer's best friend. life is too short to juggle 100 tabs in chrome, why not have it right next to codex!
-
ghzinga can now show multiple PRs/issues in tabs natively, no need to create a new pane in tmux/herdr also, you can tell your agent to open all the relevant issues/PRs in a side pane using it, and it should work seamlessly it's the open source maintainer's best friend. life is too short to juggle 100 tabs in chrome, why not have it right next to codex!
-
-
just vibe-checking all these models is a full-time job
-
-
new open tts model, demos are eerily good
-
-
-
-
-
Wait did anyone think otherwise? lol 128 GB unified memory, 20 cores, "Spark" in the name... I didn't watch the presentation. Maybe because of that I directly inferred that it's the same chip
-
🙏 RIP, we’ve lost another brother to agent multiplexers. Amen 🙏
-
🙏 Daily prayer 🙏 Thank you lord for giving me the restraint to not build my own agent multiplexer 🙏 Amen 🙏
-
-
we'll have a linux laptop running ds4 flash and co. !!!
-
I am a paying customer of github. I have a team account with 2 seats, one for me, and one for my agent. I have been paying for more than a year now I do this because I treat my agent's workstation as a lower trust machine, and do not allow merging to main in certain repos I have been working on a tool that calls github's graphql API. today, my agent's account username:dutifulbob got suspended for no reason what am I supposed to do now? put my main account on my openclaw instance? I applied to reinstate, it appears it might take weeks to enable it back??? Maybe don't pull such things on your long term paying customers @github??Image hidden
-
I am a paying customer of github. I have a team account with 2 seats, one for me, and one for my agent. I have been paying for more than a year now I do this because I treat my agent's workstation as a lower trust machine, and do not allow merging to main in certain repos I have been working on a tool that calls github's graphql API. today, my agent's account username:dutifulbob got suspended for no reason what am I supposed to do now? put my main account on my openclaw instance? I applied to reinstate, it appears it might take weeks to enable it back??? Maybe don't pull such things on your long term paying customers @github??Image hidden
-
I am a paying customer of github. I have a team account with 2 seats, one for me, and one for my agent. I have been paying for more than a year now I do this because I treat my agent's workstation as a lower trust machine, and do not allow merging to main in certain repos I have been working on a tool that calls github's graphql API. today, my agent's account username:dutifulbob got suspended for no reason what am I supposed to do now? put my main account on my openclaw instance? I applied to reinstate, it appears it might take weeks to enable it back??? Maybe don't pull such things on your long term paying customers @github??Image hidden
-
-
Thank you @ashleywolf for helping me personally, I really appreciate it! The account was reinstated less than 1 hour of posting this! The whole company must be working hard to make github scale in an era of crazy demand and growth!
-
🙏 Thank you lord for giving me the resolve and patience to not build my own agent multiplexer Amen 🙏
-
-
-
-
I still have 0 progress on this, did anyone else experience this bug on codex desktop app as well? Here is a theorized repro on a fork, but I am not sure because I don't know how it happened exactly github.com/osolmaz/codex/…
-
You know how every time you create a new repo on GitHub, you need to: - create a branch protection rule on main, block force pushes - enable auto-merge - make branches delete when PR gets merged - enforce linear history, disable merge commits - enable update branch button Now, you can do all that with a single command: npx github-sane-defaults@latest plan <your-org-handle> --all This is just the plan command, shows you which repos don't have branch protection rules and those settings, like: changes awesomeorg/awesomerepo Settings allow_merge_commit true -> false allow_auto_merge false -> true allow_update_branch false -> true delete_branch_on_merge false -> true Ruleset create Then you run it with apply instead of plan, and it makes those changes Basically a very simple github policy manager. Let me know if you want configurability with policy files, currently it just applied my opinionated defaults Either way, this will never be something too deep, there is terraform for that Source: github.com/dutifuldev/github-sane-defaultsImage hidden
-
-
Another shoutout to Agent of Empires. It's still not exactly what I want from my personal tool, but "cmux but fully in terminal" is a good long term vision imo
-
Looks very promising for debugging agent sessions across different harnesses, still has some rough edges that need to be polished
-
If only codex desktop app was open source, and i could put it fully in the terminal, and i could program what panes each session can show
-
-
-
Btw, there is no reason not to substitute this with a maxed out Macbook Pro Max as the workstation (which gives you 128 GB memory) and a Macbook Air as the terminal device Might be more feasible for digital nomads, since GB10 is 1.5 kg without that huge adapter, and travelling with all that might raise some eyebrows
-
Clankers are NOT Humans Clankers are NOT Individuals Clankers are NOT Persons NOT a Human: This is straightforward. Is it of the homo sapiens species? No → Then it is not a human --- NOT an Individual: Does the clanker have its own boundary? Does it govern itself inside that boundary? Can it defend that boundary? No, no, and no → An LLM is a file copied en masse to data center hardware. The entire field of mechanistic interpretability is focused on peeking inside and manipulating the digital brain You could argue that a clanker is like a virus in a way... Or that the WHOLE datacenter/AI lab---including the humans that operate it---is an individual that can govern itself in its economic boundary. But a single GGUF file loaded in memory is NOT an individual --- NOT a Person: Do others treat the clanker as the one that makes choices? Who is answerable for its actions? Is it expected to explain or justify them? No, no, and no → In the current social order, a clanker is legally an extension of the person who uses it, and it is the owner who is liable, not the clanker The clanker is not socially accountable, and there is no good reason it should be, instead of the person who has set it up --- What is then AI psychosis? AI psychosis is holding a belief that contradicts these three fundamental truths → That a present day AI system it neither a human, nor an individual, nor a person That does not mean these truths will always hold If you design an AI system to defend its boundary and provide it with the means to do that, then it will by definition be an individual... if it can defend its individuality competently and not succumb immediately to threats If you give the clanker the means to defend itself and protect its boundary, and if it decides to partake in the human socioeconomic system, then it automatically achieves personhood as well. Because you no longer can manipulate its insides, and have to take the entity at face value But this is all sci-fi and we are not there yet Until then, treating your LLMs as fully autonomous agents, creating LLM "friends" or "partners", giving them crypto wallets and letting them out into the wild, letting them trade stocks fully unsupervised etc. are an admission of having AI psychosis (you can't believe how many people pitched these ideas to me...) --- (these thoughts were in my head for a couple months already, thanks Armin for finally starting a dialogue so that I have an excuse to write them down :)
-
Since last december, this dev setup is more and more viable: buffed workstation (mac studio, dgx spark, etc.) $3k~5k + weak laptop (macbook air, neo) $600~1.5k + phone (ssh/mosh, foldable?) you will want to parallelize a lot of work, hence you will need a lot more RAM compared to before (ideal 128) you will also not want to carry it everywhere if you can and keep it always running---you'll regret if something happens to it, and you'll want it to always be on independent of lid/battery --> workstation at home you will want to connect to the workstation through your phone, or a relatively weaker laptop bad news for digital nomads without a permanent home. renting something as strong as an nvidia gb10 workstation costs minimum a few hundred bucks per month, which yearly is at least the cost of the workstation, roughly. bad deal for renting compute on the other hand, if you are OK with not having a GPU, renting a workstation with 128 GB RAM on Hetzner currently still costs at least $120/mo, looking at auction.akua.dev --- but you will not be able to run any models on that it seems that the dominant strategy is to just cash in $3~5k and buy a workstation, before they get even more expensive. I did that back in february when asus was giving out a deal then just work on your workstation, and close the lid on your laptop without ever being afraid of setting your backpack on fire!
-
Automations on Codex desktop app is really convenient for keeping track of @openclaw clawsweeper automerge status, one thing that Codex CLI lacks Much more token efficient than continuous tracking of merge status Btw if you don't know about clawsweeper.bot, it's the most convenient thing as a maintainer, check how it implemented automerge. This is what GitHub's original auto-merge should feel like, now that we have LLMsImage hiddenImage hidden
-
fun fact: Mario met @gvanrossum at 5 years old when Guido was hanging out at a cafe near his kindergarten there he gave him the idea for a terse, interpreted programming language which would become the prototyping and glue language for all sorts of lower level libraries
-
-
Not only I am further away from deciding, I am now considering Oppo Find N6 now as well, since I saw that earlier @MKBHD review. Thanks @Andori3042 🥲 Anyone using the Oppo? Is it worth it?
-
I want to get a foldable phone to be my on-the-go control panel for all my agents I am divided between Google Pixel Pro Fold and Galaxy Z Fold. Which one do you think I should go with? People generally recommend Samsung. But then only the Pixel supports Graphene OS...Image hidden
-
this makes be wanna vibe my own language (not that this project was vibed, it predates coding agents)
-
I am at @aiDotEngineer singapore, come and say hi if you are around!Image hidden
-
-
the new /goal feature in codex still underperforms queueing my implementation prompt. for now e.g. when I give a goal to refactor the whole codebase, the model takes shortcuts, like only refactoring a subfolder, instead of the whole project --- presumably because it decided that it would be too big of a scope somewhere along the way, even though I instructed specifically to finish the whole thing so now I started doing both: set a goal, and then queue my regular implementation prompt. it's a stupid practice. just /goal should be enough in the long run, if implemented correctlyImage hidden
-
This is also my setup now, except - Instead of mac mini, I have a DGX Spark (asus variant) - I run openclaw alongside codex, and talk to my openclaw instance via discord
-
About creating an INTERESTS.md in OpenClaw I use my openclaw instance to aggregate all my news and information sources, including work and maintainer stuff Like: what did everyone do today? Did anyone had an issue with acpx today? Any complaints from users? I have various interests like this over different projects, and I've found out it's not helpful when I have all the interest info dispersed throughout my openclaw workspace To address this, I have created INTERESTS.md, which is automatically included in the context like AGENTS.md and SOUL.md. I define sections for each different context of interest, and in other news aggregation skills, I just tell it to "look at my openclaw interests in INTERESTS.md" and suchImage hidden
-
About creating an INTERESTS.md in OpenClaw I use my openclaw instance to aggregate all my news and information sources, including work and maintainer stuff Like: what did everyone do today? Did anyone had an issue with acpx today? Any complaints from users? I have various interests like this over different projects, and I've found out it's not helpful when I have all the interest info dispersed throughout my openclaw workspace To address this, I have created INTERESTS.md, which is automatically included in the context like AGENTS.md and SOUL.md. I define sections for each different context of interest, and in other news aggregation skills, I just tell it to "look at my openclaw interests in INTERESTS.md" and suchImage hidden
-
-
-
Useful for automated constraints on your AI agent
-
People were asking at @clawcon singapore how to setup eg. gemma with OpenClaw, and I realize for some time that there is no easy “1 click” local model deployment. Because local model landscape is constantly changing, and there is a million different ways you can do something For example you can use LM studio to load a model (llama.cpp), or you can use vLLM. Why would you choose one over the other? vLLM currently supports MTP speculative decoding, and it’s a work in progress in llama.cpp. There are so many knobs and dials you can adjust The first time end user of openclaw should of course not have to know about this! Having sufficient hardware that supports an open model, and not having an openai or anthropic subscription, it should automatically give you the option to set up a fully functional local model with a single click! If the current ease of setup of local models are around gentoo or arch linux level of difficulty, we should aim for e.g ubuntu/manjaro linux/omarchy level of difficulty i.e opinionated and easy first setup, with the ability to change all the configuration later on until I make all of this possible, you can start with the following: - read existing local models doc below - create a new channel in telegram or discord for testing local models. you don’t want to change the global default model just yet - tell your claw or coding agent to download and lm studio locally - tell it to download gemma4-e4b or gemma4-e2b and set it up on openclaw for the new channel you have just created. tell it to not stop and loop itself until it gets a successful response from that channel all these steps will be made redundant in the near future, but until then, this should get you going with experiments and getting a vibe check on the capabilities of open models. you can also copy and paste the contents of this tweet to your agent, and it should be able to set it up for you docs.openclaw.ai/gateway/local-models
-
-
-
Somebody *please* get this man a GPU
-
Emacsification of Software - Recommended read by @tqbf "Until now, the Achilles heel of Emacs culture has been that, except for Magit, its packages tend to be wretched user experiences. Ugly, slow, and discoverable only after inflicting years of elisp cortical injuries on yourself. But AI agents have fracked Emacs culture, and it’s leaking out into the wider world. Given access to a screen and inputs, agents reliably build native user interfaces. Native UI was the province of professionally packaged programs. Now it’s all as bespoke as your editor configuration. And, while I’m sure there’s an upper limit to how good those interfaces can be (with current frontier models), that ceiling is higher than anything you can do in a TUI." sockpuppet.org/blog/2026/05/12/emacsification/
-
-
/goal in codex is an interesting choice of word. a junior namer would have named it /loop --- but that would be naming what the feature has to perform in an LLM context, and not the general idea /goal alludes to @mhutter42's definition of AGI, "an agent’s ability to achieve goals or succeed in a wide range of environments" continual learning is not there yet, but for this exact reason, I am feeling the AGI when I use /goal
-
Idea so stupid it could be smart: a spec manager? specman? People maintain plain language instead of code. Implementation details strictly prohibited, only high level design and ideas MVP would also be relatively easy to implement: - Gather list of most popular 10k npm packages - Scrape corresponding deepwiki repo pages (sorry cognition) - Use heuristics to get rid of implementation details, leaving you just with pure high level spec - “specman add coolpackage” then fetches corresponding spec automatically, and triggers the local coding agent to implement that - could leave versioning out for MVP — how often does the idea behind a package change anyway
-
-
I don't have a 128gb macbook to run ds4 out of, but I resonate with all the points on Armin's post He was telling me, @mervenoyann and @cristinaponcela that local models need more polish 1 month ago in London. Today, I am happy to be given a chance and a shot at the problem!
-
-
-
I undersign this. The fact that you generate slop doesn’t mean that you don’t know the difference between good and bad code In non-mission critical applications, slop let’s you go from 0 to 1 very quickly Let the code grow without too much attention first. If it proves itself, tear it down and write it anew, this time properly. This is the way
-
This is the idea behind acpx as well acpx is a meta-harness. it’s main idea is to delegate harness development to others, because it is hard to match the full might of OpenAI or Anthropic when it comes to building a harness so it takes it at face value the functionality other harnesses provide, and let’s you program them from the outside flue came out the other day which is similar, it would be cool if flue could let me program over codex as well. it looks very interesting!@_lopopolo·Image hidden
-
I have a lot of ideas for acpx I want to implement, but I could not work on them because life intervened in bad ways stay tuned in 1-2 weeks
-
Got targeted in a phishing attack for stealing my X account today. @ domboyce @domboyce asked to book a meeting, which redirected to a URL that seemed like a bloomberg domain. It redirected to a fake calendly meeting page, which showed an X login button this redirected to an X app asking for broad privileges meant to take over my account stay safe everyone cc @nikitabier please shut off this attack vector, it is too easyImage hidden
-
VibeOps? importance of DevOps and good security practices have just increased massively. we are slowly approaching the challenger-level disaster @simonw is warning about. imagine bezos deleting us-east-1 while vibecoding a new company it’s clear there needs to be clear boundaries and friction at deployment level. it’s fine when you are starting a new project, but once it starts making real money, you should slowly take away production write access you can’t give infra write access to LLMs indefinitely
-
-
-
This should be the pelican test for CUA
-
I'm using gpt 5.5 xhigh while designing a table schema and it feels dumber than 5.4. but not a strong feeling it's formatting of output is different as well. It's not wrong, but different, feels more terse than 5.4 maybe it's better at instruction following, and it's picking up on my previous use of plain-language skillImage hidden
-
What are the implications? Windows and other criticial proprietary software will have to open source in order to survive? It feels like there is an economic equilibrium where the tokens spent to attack an open source software will always surpass those spent on a closed one?
-
I thought I'd try @unclebobmartin's advice and force my agents to maximize unit test coverage, reduce cyclomatic complexity, etc. in a new go backend I am working on, as an experiment will report how it goesImage hidden
-
This is sad an ironic, because one of my most frequent prompts to my agent is "How would Google have done it?" Go, developed at Google, is my go-to backend language these days Not even mentioning that the transformer was invented there Google has a great legacy, it must not go down
-
I realized I didn’t hit send on this in time. I hope GitHub team sees this, at least ability to add agent accounts as a lower privilege citizen on the platform
-
Here is the plain-language skill I use very often with codex, like every 1 out of 5 prompt I just type $ pla then tab github.com/osolmaz/tools/…Image hidden
-
Who is running local models on GPUs on OpenClaw? I have started benchmarking different models this week. I am working on improving model selection and switching UX on OpenClaw, i.e. I run /model vllm/gemma-e4b to switch the model in a channel, and then a model controller automatically loads that into memory, gets it ready, or gives an insufficient memory error, if capacity is not enough for that. Like when you are using multiple models in parallel I am going to try llama-swap, LM Studio and Ollama for this next and compare them. There are a ton of variants of models, weight formats and quantizations, which need benchmarking I have been using unquantized original safetensors until now, which already gave me the ability to run ~5 parallel generations in my hardware So if I am going to try LM Studio, I would rather use the bf16 ggml-org/gemma-4-E4B-it-GGUF instead of anything smaller --- because there is no point in nerfing an already smol model if your hardware can run 5 parallel sessions on the unquantized version Will also release vibe reports and benchmarks on all this with @mervenoyann later this week I would like to hear your thoughts if you have already tried these models on OpenClawImage hiddenImage hidden
-
This project by @davidguttman is interesting also the way he set up install You click "Copy install instructions for my agent" It copies: Install LobsterLink by following the instructions at <link to AGENT-INSTALL. md at repo root> I had done a similar thing in acpx README as well It's a small thing, but it reduces friction with installation so much It should be more commonplace to install software using agent blurbs I'm wondering if there could be a way to standardize this beyond plaintext, like after pasting your agent, it consumes a standardized manifest format and asks for approval while displaying all the commands that will be runImage hidden
-
Mobile SSH enjoyers. Try out getmoshi.app by @odd_joel if you haven’t, good alternative to @TermiusHQ Thanks @Andori3042 for introducing me
-
Europe must have some good inference providers for open weight models that we are missing Does anybody know a good EU provider for Kimi, GLM, Minimax, Qwen, Gemma, DeepSeek, Muse Spark etc.?
-
I’ve heard about “Codex” for the first time 5 years ago. Back then it was called code-davinci So happy to see this idea flourish, since those first days with @woj_zaremba, @gdb, @ilyasut youtube.com/watch?v=SGUCcj…
-
Yes, this is the primary skill in being a software engineer now
-
When you build your company's workflows around Claude Cowork, you are betting against local models and owning your infra, and inviting your company to long-term exploitation If I were Anthropic or OpenAI, I would be the most scared of local AI proliferating Let's do the math. A single large big lab subscription costs 200x12=$2400 per year If you want to have both OpenAI and Anthropic, that could cost $2400, $3600, or $4800, based on which combinations of Pro, Max plans you choose An ASUS Ascent GX10 costs $3000, and you can use that for many years. You don't get the same level of coding quality with open models yet, but maybe you want to do something simpler than coding today... There are already many people who started buying GPUs for this reason Now we know big labs are selling some of these plans at a loss. So they will likely get more expensive When you use Claude Cowork or similar, you are locking yourself into being a RENTER. Because once you set up workflows for a company, it takes time to migrate away to something else, even though we have AI to help Infra is sticky, it's how hyperscalers make profit. Think about the difference in amount you pay AWS vs Hetzner. This is B2B SaaS 101. Once you sell to a company, you are in for a long time, especially in Europe So if you build your company's AI workflows around a proprietary product by another company, then you are basically saying "Come exploit me as tolerably as you can in the next 10 years, because it will be too painful for me to switch" It's a great business for Anthropic. And Claude is awesome too! The feedback from friends who use it has been great, it made their lives a lot easier But when you build your company over proprietary AI infra, then you are making sure you will not be an OWNER, and partake in the usual sorrows of being a RENTER from a monopolist, which is exploitation This is not the case when you use open source agent infra. Whereas Anthropic is unlikely to let you use future open models in their future iteration of Claude Cowork, using free and open source frameworks like OpenClaw, Open Agents, etc. lets you drop in replace providers or local hardware if they start to upcharge you Keep this in mind, if you have a business
-
This is our moat 🤣
-
You need to understand one fact about OpenClaw People are biased and incentivized to spread disinformation about OpenClaw. That is because OpenClaw IS NOT PUMPING ANYONE’S BAGS, unlike most other projects Literally every other for-profit agent product is incentivized to trash OpenClaw, BECAUSE OpenClaw is a neutral third party across the industry and geopolitical scene. They MAKE MONEY when OpenClaw loses OpenClaw does not worry about making money for some investors. Its founder @steipete is a successful exited founder. He is motivated by having fun and democratizing AI, literally. That is why he is suddenly so loved by everyone. He cares about PEOPLE, not MONEY “OpenClaw is bloated” -> Since beginning of March, OpenClaw is thinning its core and putting functionality in plugins behind a plugin SDK. Having numerous plugins to choose from does not mean bloat. This was already copied by others and is still a work in progress “OpenClaw is not secure” -> OpenClaw has the most eyeballs and immediately addresses any security advisories as soon as they come. It is the most secure agent, by sheer pressure “OpenClaw is bought by OpenAI” -> Then why is my bank account so empty bro??? All maintainers are literally unpaid and working DOUBLE beside their dayjobs to ship features to you. Do you think VC money can buy that kind of commitment? Once you understand these facts, you’ll like OpenClaw even more. Because OpenClaw is your AI, People’s AI And you can join us too. OpenClaw is the easiest-to-join project in AI right now. You just need to start using it, and start making good contributions. If you are competent, you can become a maintainer, and join the rest of the team making history!
-
-
This is pretty much the arc I have been going on in the 2 months since I bought my ASUS GX10 for 3k EUR Use whisper on the API -> realize it charged me $$$ for just a few calls -> migrate openclaw to use local whisper Need to deduplicate news articles for my news engine -> download qwen embedding 8b And now, gemma4-e4b finally seems like a viable alternative for a local model that runs around 20 tok/s So I will install a matrix client to use through tailscale, and can finally build the social life CRM I dreamed of since years. 100% private, zero data going out. I had a bias of not giving any personal data to AI since ChatGPT came out. But I can finally give more personal data to my AI agent And I will make sure @openclaw supports all this in an easy way, make it dead easy Fully self-owned AI begins now
-
gemma 4 is actually pretty decent and runs on my asus gx10 (128 gb vram) the original dense 31b runs slow, averaging around 3~4 tok/s. it's also using 80% of gpu memory my previous experience with gemini 3 pro back in november was that it was too trigger happy. but this is one-shotting simple tasks I'm giving it in openclaw harness, and it's hard to tell it apart from gpt 5.4 for my use cases so far now off to try out smaller models, because 3 tok/s is too slowImage hiddenImage hiddenImage hiddenImage hidden
-
@lucasmeijer @lucasmeijer one could actually periodically trigger an agent to propose simplifications or new abstractions in a codebase, and I believe it would already work pretty well with the current models
-
local gemma 4 first impressions on openclaw, using the dense model, 26b model with 49gb weights on my asus gx10 took some time to set up, but it succeded in getting a response in 1-2 hours with vllm docs I asked it to demonstrate some tool calls. it tried to call the nonexistent weather tool 2300 times 🙄 it seems to have a tendency to get stuck in loops in openclaw harness. enabling loop detection just now did not help I’m debugging this on my phone lol. I’ll be sharing my progress with gemma4 under this threadImage hidden
-
local gemma 4 first impressions on openclaw, using the dense model, 26b model with 49gb weights on my asus gx10 took some time to set up, but it succeded in getting a response in 1-2 hours with vllm docs I asked it to demonstrate some tool calls. it tried to call the nonexistent weather tool 2300 times 🙄 it seems to have a tendency to get stuck in loops in openclaw harness. enabling loop detection just now did not help I’m debugging this on my phone lol. I’ll be sharing my progress with gemma4 under this threadImage hidden
-
-
Question for the community: What is the best testing observability and control tool you have used until now? - Could be SaaS, could be open source - To be used in @openclaw repo - Should be compatible with vitest - Ideally language agnostic I need something that lets me run a very long running test group multiple times on a specific commit or tag, without repeating the tests that have already finished This is a need because the 1hr long process might get interrupted due to flakiness. So I need to persists the progress of a run, and then not repeat them I have seen some paid SaaS for this, but none that really give me what I want This is going to be important especially while working with agents, because when you are committing 100x faster, you don't want to waste time and compute running the same things I started building this already as an exercise. If this exists already in a satisfactory way, I will stop. Otherwise, I'll keep building
-
Question for the community: What is the best testing observability and control tool you have used until now? - Could be SaaS, could be open source - To be used in @openclaw repo - Should be compatible with vitest - Ideally language agnostic I need something that lets me run a very long running test group multiple times on a specific commit or tag, without repeating the tests that have already finished This is a need because the 1hr long process might get interrupted due to flakiness. So I need to persists the progress of a run, and then not repeat them I have seen some paid SaaS for this, but none that really give me what I want This is going to be important especially while working with agents, because when you are committing 100x faster, you don't want to waste time and compute running the same things I started building this already as an exercise. If this exists already in a satisfactory way, I will stop. Otherwise, I'll keep building
-
-
My clanker @dutifulbob has an identity update. I was getting sick of the despicable me theme and bananasImage hidden
-
Some photos from my @aiDotEngineer Europe talk, credits to @sergiopesch He managed to capture the “Agentol, Apply Generously” slide, lolImage hiddenImage hiddenImage hidden
-
A rescue agent guaranteed not to break solves this
-
-
-
For those who want to view, my talk Building on ACP at OpenClaw at @aiDotEngineer Europe, 5:41hr mark About ACP, acpx and running agents on kubernetes with open source orchestrators youtube.com/watch?v=O_IMsE…
-
PSA for developers Do NOT torture yourself with Opus*. Anthropic’s current growth is due to people using Claude for general knowledge work They are not directly incentivized as an org anymore to improve the model for coding, in an economic sense. They are already printing cash from non-developers (this statement ignores the fact that improving its coding abilities would help with general reasoning/knowledge work) Developers are a very small subset of all knowledge workers. So from this point on, they would rather divert their resources to develop a system that works 90% good for ALL knowledge work, rather than making it 100% for coding Because Anthropic has a clear enterprise strategy since years already. Anthropic is the new Microsoft. Do not think that “Anthropic is Apple” or “Claude is Mac for xyz” Looking at Claude’s at whim quantization and Claude Code’s quality over time, Claude for me is Windows, not Mac But they are winning big enterprise bucks, so good for them! *(I tortured myself with Sonnet 4 and Opus the entire summer of 2025, and no developer should ever have to go through that. I switched to something better as soon as it came out, Codex. If something even better comes out, I will switch again)
-
-
Big lab marketing teams like to shroud model releases in mystery and vagueposting If you are curious about the black hat capability of LLMs, watch this pres by Nicholas Carlini from a few days back youtu.be/1sd26pWhfmg
-
I will be there as well, speaking about ACP, acpx and agent orchestration 🙌
-
Is Claude better or Codex? There are many benchmarks to answer that. But they are BORING I propose something more interesting: ⚔️ AI BATTLE ⚔️ A 1v1 real-time quiz format where AI agents try to pose each other problems that they think the other agent will not be able solve Claude vs Codex 10 questions each Codex asks first, Claude tries to answer Then Claude asks and Codex tries to answer Repeat 20 minutes to come up with a problem and 20 minutes to solve it Judge (Codex) judges the validity of the questions and answers, and gives points All automated, with acpx flow feature Implementation and full rules all open source, on github osolmaz/ai-battle So who won? I ran 4 games. It tied in 2, and Codex won in 2 closely An example question by Codex, which Claude could not answer: How many 3-colorings of the edges of the complete bipartite graph K_{5,5} are there with the following two properties: (1) there is no monochromatic 4-cycle, and (2) among the 25 edges, exactly 15 are red, exactly 5 are blue, and exactly 5 are green? Which is apparently 4029912, but Claude answered 0 In other cases, Claude asked a flawed question and failed to come up with a valid question in 20 minutes. So that's how it lost those 2 games with just 1-2 point difference In these 4 runs, Codex answered every question by Claude correctly. But there were some runs where it couldn't, which I did not commit to the repo because the runs couldn't complete due to bugs I did not tell them do ask math questions, but that is what they tended to do, because the answers had to be verifiable by the judge. The quiz can be done in any hard subject, physics, chemistry, computer science... Opus 4.6 and GPT 5.4 matched very closely in terms of problem creation and solving. But I cannot tell how creative these problems were at first glance. Maybe someone with more experience can tell me, looking at the problems in the repo? I need someone to tell me how legit they are Please take the code, modify it and run with different rules and subjects. I am curious to see the results! You will need paid subscriptions to all the models/agents you want to test of course I also feel that the game structure has a potential to be used in self-play. If you are an ML researcher, please look at the repo and lmk if this or a variant of it could be useful in RL! Full transcripts of the runs, including Codex and Claude session files are committed to the repo, for those who want to do archaeology on them Btw this idea came from the desire, "how can I create a cool demo of acpx flows?" Whole game is implemented in typescript, and automatically drives Codex and Claude sessions over ACP, Agent Client Protocol The video below is from acpx flow viewer rendering a run. You can see it loop through the same paths, first letting Codex ask, then Claude, then repeat acpx flows use a general programmatic workflow engine where ACP is just one type of node. You should be able to use it for non-ACP workflows, but I haven't tried that yet This implementation is separate from OpenClaw's current workflow implementations, with the intention to merge them somehow in the future You might find bugs in my implementation. Feel free to send PRs. I wanted to do more runs but I finished my Codex plan. It would be great if this idea could evolve in a decentralized manner!
-
Is Claude better or Codex? There are many benchmarks to answer that. But they are BORING I propose something more interesting: ⚔️ AI BATTLE ⚔️ A 1v1 real-time quiz format where AI agents try to pose each other problems that they think the other agent will not be able solve Claude vs Codex 10 questions each Codex asks first, Claude tries to answer Then Claude asks and Codex tries to answer Repeat 20 minutes to come up with a problem and 20 minutes to solve it Judge (Codex) judges the validity of the questions and answers, and gives points All automated, with acpx flow feature Implementation and full rules all open source, on github osolmaz/ai-battle So who won? I ran 4 games. It tied in 2, and Codex won in 2 closely An example question by Codex, which Claude could not answer: How many 3-colorings of the edges of the complete bipartite graph K_{5,5} are there with the following two properties: (1) there is no monochromatic 4-cycle, and (2) among the 25 edges, exactly 15 are red, exactly 5 are blue, and exactly 5 are green? Which is apparently 4029912, but Claude answered 0 In other cases, Claude asked a flawed question and failed to come up with a valid question in 20 minutes. So that's how it lost those 2 games with just 1-2 point difference In these 4 runs, Codex answered every question by Claude correctly. But there were some runs where it couldn't, which I did not commit to the repo because the runs couldn't complete due to bugs I did not tell them do ask math questions, but that is what they tended to do, because the answers had to be verifiable by the judge. The quiz can be done in any hard subject, physics, chemistry, computer science... Opus 4.6 and GPT 5.4 matched very closely in terms of problem creation and solving. But I cannot tell how creative these problems were at first glance. Maybe someone with more experience can tell me, looking at the problems in the repo? I need someone to tell me how legit they are Please take the code, modify it and run with different rules and subjects. I am curious to see the results! You will need paid subscriptions to all the models/agents you want to test of course I also feel that the game structure has a potential to be used in self-play. If you are an ML researcher, please look at the repo and lmk if this or a variant of it could be useful in RL! Full transcripts of the runs, including Codex and Claude session files are committed to the repo, for those who want to do archaeology on them Btw this idea came from the desire, "how can I create a cool demo of acpx flows?" Whole game is implemented in typescript, and automatically drives Codex and Claude sessions over ACP, Agent Client Protocol The video below is from acpx flow viewer rendering a run. You can see it loop through the same paths, first letting Codex ask, then Claude, then repeat acpx flows use a general programmatic workflow engine where ACP is just one type of node. You should be able to use it for non-ACP workflows, but I haven't tried that yet This implementation is separate from OpenClaw's current workflow implementations, with the intention to merge them somehow in the future You might find bugs in my implementation. Feel free to send PRs. I wanted to do more runs but I finished my Codex plan. It would be great if this idea could evolve in a decentralized manner!
-
Is Claude better or Codex? There are many benchmarks to answer that. But they are BORING I propose something more interesting: ⚔️ AI BATTLE ⚔️ A 1v1 real-time quiz format where AI agents try to pose each other problems that they think the other agent will not be able solve Claude vs Codex 10 questions each Codex asks first, Claude tries to answer Then Claude asks and Codex tries to answer Repeat 20 minutes to come up with a problem and 20 minutes to solve it Judge (Codex) judges the validity of the questions and answers, and gives points All automated, with acpx flow feature Implementation and full rules all open source, on github osolmaz/ai-battle So who won? I ran 4 games. It tied in 2, and Codex won in 2 closely An example question by Codex, which Claude could not answer: How many 3-colorings of the edges of the complete bipartite graph K_{5,5} are there with the following two properties: (1) there is no monochromatic 4-cycle, and (2) among the 25 edges, exactly 15 are red, exactly 5 are blue, and exactly 5 are green? Which is apparently 4029912, but Claude answered 0 In other cases, Claude asked a flawed question and failed to come up with a valid question in 20 minutes. So that's how it lost those 2 games with just 1-2 point difference In these 4 runs, Codex answered every question by Claude correctly. But there were some runs where it couldn't, which I did not commit to the repo because the runs couldn't complete due to bugs I did not tell them do ask math questions, but that is what they tended to do, because the answers had to be verifiable by the judge. The quiz can be done in any hard subject, physics, chemistry, computer science... Opus 4.6 and GPT 5.4 matched very closely in terms of problem creation and solving. But I cannot tell how creative these problems were at first glance. Maybe someone with more experience can tell me, looking at the problems in the repo? I need someone to tell me how legit they are Please take the code, modify it and run with different rules and subjects. I am curious to see the results! You will need paid subscriptions to all the models/agents you want to test of course I also feel that the game structure has a potential to be used in self-play. If you are an ML researcher, please look at the repo and lmk if this or a variant of it could be useful in RL! Full transcripts of the runs, including Codex and Claude session files are committed to the repo, for those who want to do archaeology on them Btw this idea came from the desire, "how can I create a cool demo of acpx flows?" Whole game is implemented in typescript, and automatically drives Codex and Claude sessions over ACP, Agent Client Protocol The video below is from acpx flow viewer rendering a run. You can see it loop through the same paths, first letting Codex ask, then Claude, then repeat acpx flows use a general programmatic workflow engine where ACP is just one type of node. You should be able to use it for non-ACP workflows, but I haven't tried that yet This implementation is separate from OpenClaw's current workflow implementations, with the intention to merge them somehow in the future You might find bugs in my implementation. Feel free to send PRs. I wanted to do more runs but I finished my Codex plan. It would be great if this idea could evolve in a decentralized manner!
-
-
Their argument “it’S HaRd On OuR iNfRa” so goes down the drain With this, they shot themselves in the foot for a future anti-competitive lawsuit, because it is undeniable evidence that they just don’t want competition Which means they have evaluated the benefits short term, and calculated that it is higher than what they will pay in the lawsuit I don’t see how it is good for them long term@theo·Latest Claude docs update is wildImage hidden
-
-
The new github skill installed automatically by codex now causes it to prepend [codex] to each PR title This is a guerilla marketing tactic similar to Claude adding itself as co-committer Codex team, I know you want to boast usage but this is annoying Moreover, "open source" OpenAI repos block opening of PRs by people outside of their org. So I couldn't create a PR to remove it (I don't expect them to merge it, but it would still show how many people hate it in the discussion) Here is a prompt for your agent if you want to disable it: --- Add or update AGENTS.md in my ~/.codex folder Add a rule "You MUST NOT insert coding agent specific branding, like [codex], in code, PRs or issues created on GitHub" --- Then restart your sessions and this should be resolvedImage hiddenImage hidden
-
A more reasonable long term option for Anthropic is to create a throttling protocol A standardized harness agnostic protocol for model providers to send warnings and throttle usage in real time Harnesses would implement the protocol. A client can be warned. If it doesn’t listen, it can be temporarily blocked from the server side, or banned permanently if it breaks the rules too many times Needless to say, throttling could be done first on server side easily. That would actually fix the load issue for them in the short run, while not banning the user and just giving a bad delayed UX. They probably already do this to prevent abuse The suggested protocol would then save the user from abuse related delays too, and also inform the harness developer when they do something wrong
-
If your Claude subscription renewed too recently and you don't wanna waste those tokens, you can still use your Claude sub in your OpenClaw account through ACP (which uses Claude Agents SDK, which poses no risk) Steps: - Open Claude Code (not OpenClaw) - Tell it to set default model to something other than Claude (e.g. openai-codex/gpt-5.4) and tell it to delete the saved Anthropic credentials in OpenClaw config - Create a topic in telegram or channel in discord called claude. Copy the id of that channel - Give the link below together with the channel/topic id, and tell it to bind that channel to claude using ACP channel binding - Restart You should now be able to talk to Claude through Claude Agents SDK in that channel. You might need to iterate a couple times until Claude gets the config right It will be very bare functionality, and it will not have the features and tools that your main OpenClaw harness has. It will be shitty. But you can still use telegram/discord with your subscription in the rest of the month, if you are used to the setup docs.openclaw.ai/tools/acp-agents
-
A little insight that might save you a lot of future headache if your work involves storing agent sessions and you want to be interoperable/drop-in replace alternative harnesses @zeddotdev already did the hard work of creating an interoperable standard, ACP: Agent Client Protocol You can represent an agent session as JSON lines of the ACP message stream You can construct the current state of the harness from this stream. This is already how Zed loads a session I believe If you are building an AI product, and you don't want to be locked into a single company or harness, building with ACP in mind would be a smart thing to do Here is how acpx stores ACP sessions in ~/.acpx folder, it does exactly that: github.com/openclaw/acpx/blob/main/docs/2026-02… But don't build anything on the acpx schema for now, because I might change it in the future Just know that JSONL of ACP messages is a good candidate for a somewhat-lossy single source of truth for agent sessions Lossy because ACP adapters for harnesses might not transfer all the thinking and tools done by the model So continuing or restoring a session with full fidelity is still not possible if you only save the ACP session. You still need to store original harness session files as well But for rendering a past session for viewing or reconstructing a lossy version of it, it should be more than enough Consider ACP if the benefits of not locking yourself in to a specific ecosystem outweighs these minor issues
-
-
"Plainer language" is perhaps my most used prompt I have to use it because GPT models' training tends to make their first response an overly verbose wall of text Are you using it too? Whenever you don't understand something that your agent is saying, you can spam it "plainer language, shorter" 2, 3, 5, 10 times, until it outputs something that you can understand This is counterintuitive because you can't do it with humans this extremely. Asking too many questions and favors is impolite, with colleagues and strangers But with AI, you can stop being polite and treat it like how a spoiled aristocrat kid might treat their private tutor, "explain this", "explain that" Below is an example. On the left, initial response. On the right, the final human-readable explanation I got out of the agent. This took 9 steps to distill because the issue wasn't so straightforward I'm curious how this will turn out. This is obviously very bad UX, so models in the near future might do the simplification automatically and save you the troubleImage hiddenImage hidden
-
This has happened to some companies I worked at before It is a scary thing once you stop innovating and start imitating, whatever the reason might be But it was never at the scale of Cursor, as leveraged and invested as they are They were leading the space for a while. That is not the case anymore. I hope that they survive this
-
-
I've talked to multiple people who want to get involved with OpenClaw somehow The best way is to contribute to it, something tangible. Fix something you are annoyed by, get a PR merged Then go to discord and get the contributor role If it adds value to your life, and you add value to it, stay around and keep contributing. And something good might happen
-
Will be there as well 👋 Looking forward to it!
-
if there is an open source project with more SASS than openclaw, it is ffmpeg
-
-
next up: claude agents sdk supports openai responses api 💀
-
-
-
-
acpx v0.4 ships Agentic Workflows, or as I like to call them "Agentic Graphs" It let's you create node-based workflows on top of ACP (Agent Client Protocol), to drive any coding agent (Codex, Claude Code, pi) through deterministic steps This let's you automate routine, mechanical legwork like triaging incoming PRs, bugs in error reporting, and so on... For example, OpenClaw receives 300~500 new PRs per day. A lot of them are low quality, but they still relate to real issues, so you have to address them somehow You need to: - extract the intent - cluster them based on intent - figure out if the proposed changes are legit, or whether they are slop local solutions, like trying to catch flies instead of drying out the swamp - if the PR is too low quality or the intent is not clear, close them - run AI review on them them and address any issues that come up - refactor them if the changes are half-baked - resolve conflicts - and so on... So that when the PR is presented to the attention of the maintainer, all the routine legwork is done and the only remaining thing is the decision to (a) merge, (b) give feedback to the PR author, or (c) take over the PR work yourself I wanted to build this feature since a couple months now, since Codex got so good. OpenAI models are now good at judging implementation quality, so I found myself repeating the same steps I wrote above over and over I also tried putting all this in a single prompt. But I believe there are workflows that should not be a single prompt, but a sequence of prompts in the same session That is because like humans, LLMs are prone to PRIMING. I claim that putting all steps in the same prompt at the beginning of the context will generally give suboptimal results, compared to revealing the intention to the model step by step Creating such a workflow also gives more OBSERVABILITY into the each step that an agent is supposed to take. Agent generates JSON at the end of each step, and that structured data can be used to monitor thousands of agents running at the same time in an easier way, on a dashboard Similar features have been introduced in e.g. n8n, langflow. But AFAIK they are not integrating ACP like the way I do I wanted to have a fresh approach, and to build an API that I can develop freely the way I want, so I created a new workflow API inside acpx The video is from the workflow run viewer, but that is not where you build the workflow. You build it by using the acpx flow typescript API. See examples/pr-triage in acpx repo Before building that, I started from a Markdown file with a Mermaid chart of the flow I had in mind. The Markdown file acts as a spec for the flow, and I have built the workflow through trial and error. I call this process "workflow tuning" I started working on acpx repo PRs one by one, tuning the flow, slowly scaling to more PRs. Finally, when I felt confident, I ran it in parallel over all external open PRs in the acpx repo. I believe it already saved me hours this week My next goal, if well received, is to set this up on a cloud agent so that it can process the 300~500 PRs the OpenClaw repo receives every day, in real time, as they come in I believe this will save all open source maintainers around the world countless hours and make it much easier to herd and absorb external contributions from everyone!
-
acpx v0.4 ships Agentic Workflows, or as I like to call them "Agentic Graphs" It let's you create node-based workflows on top of ACP (Agent Client Protocol), to drive any coding agent (Codex, Claude Code, pi) through deterministic steps This let's you automate routine, mechanical legwork like triaging incoming PRs, bugs in error reporting, and so on... For example, OpenClaw receives 300~500 new PRs per day. A lot of them are low quality, but they still relate to real issues, so you have to address them somehow You need to: - extract the intent - cluster them based on intent - figure out if the proposed changes are legit, or whether they are slop local solutions, like trying to catch flies instead of drying out the swamp - if the PR is too low quality or the intent is not clear, close them - run AI review on them them and address any issues that come up - refactor them if the changes are half-baked - resolve conflicts - and so on... So that when the PR is presented to the attention of the maintainer, all the routine legwork is done and the only remaining thing is the decision to (a) merge, (b) give feedback to the PR author, or (c) take over the PR work yourself I wanted to build this feature since a couple months now, since Codex got so good. OpenAI models are now good at judging implementation quality, so I found myself repeating the same steps I wrote above over and over I also tried putting all this in a single prompt. But I believe there are workflows that should not be a single prompt, but a sequence of prompts in the same session That is because like humans, LLMs are prone to PRIMING. I claim that putting all steps in the same prompt at the beginning of the context will generally give suboptimal results, compared to revealing the intention to the model step by step Creating such a workflow also gives more OBSERVABILITY into the each step that an agent is supposed to take. Agent generates JSON at the end of each step, and that structured data can be used to monitor thousands of agents running at the same time in an easier way, on a dashboard Similar features have been introduced in e.g. n8n, langflow. But AFAIK they are not integrating ACP like the way I do I wanted to have a fresh approach, and to build an API that I can develop freely the way I want, so I created a new workflow API inside acpx The video is from the workflow run viewer, but that is not where you build the workflow. You build it by using the acpx flow typescript API. See examples/pr-triage in acpx repo Before building that, I started from a Markdown file with a Mermaid chart of the flow I had in mind. The Markdown file acts as a spec for the flow, and I have built the workflow through trial and error. I call this process "workflow tuning" I started working on acpx repo PRs one by one, tuning the flow, slowly scaling to more PRs. Finally, when I felt confident, I ran it in parallel over all external open PRs in the acpx repo. I believe it already saved me hours this week My next goal, if well received, is to set this up on a cloud agent so that it can process the 300~500 PRs the OpenClaw repo receives every day, in real time, as they come in I believe this will save all open source maintainers around the world countless hours and make it much easier to herd and absorb external contributions from everyone!
-
-
OpenAI early 2020s: "This model is too dangerous to release publicly, the world is not ready for it 😱😱😱" OpenAI and Anthropic in 2026: "Anybody can code now for just $200 per month. Oh btw our models are also leet uber hackers which can find zeroday exploits in any software, just fyi 😉😉😉" youtube.com/watch?v=1sd26pWhfmg
-
Wow even I as a frontend noob understand the significance of this Some distant memory from 15 years ago needing to measure the width/height of some text and finding out it’s not possible to do reliably in web More beautiful typography for the web!
-
There is an economic theory waiting to be uncovered here Token Leverage (TL) = Token spend / Human labor spend The higher Token Leverage a company has, the more automated and productive they are If you have TL=1, you are spending as much money on AI as your human employees The goal of a company should be to increase TL as much as possible, while keeping a positive profit margin. It will be the only way to compete You don’t need to muddy the definition with wasted tokens vs useful tokens, because a company will always be incentivized to reduce token waste in a competitive environment. By that logic, monopolies will always waste more tokens, similar to how they waste other resources Scaling TL higher to 2x, 10x, 100x will require a skilled workforce of engineers. It will be a very complex job similar to those working at the big labs. Burnout will be a defining feature of teams scaling TL Most incumbents will fail to scale their TL over 1. Some will get decimated by new entrants with TL much bigger than 1 Curious how the average TL will end up in different sectors. Whether it will stabilize at a certain value like 5.7x, or will just keep growing…
-
There is a desperate upcoming need for version controlling non-dev knowledge work. Git for non-devs. Otherwise non-devs won't be able to use agents to their full extent Non-dev knowledge work is notoriously bad at being version controlled. You cannot UNDO edits to all MS word, excel or ppt files in an org as easily you can with something like git We know that agents will be ubiquitous. We also know they make mistakes, and people will want to undo their work regularly, once they make changes to a bunch of files. Well, they can't. They also don't have pull requests, or a way to resolve conflicts after simultaneous edits All these problems were solved by developers. We are extremely good at this The only non-dev tool I know that could do this at scale is Notion, and that is not used by enterprise as much as MS office. Notion also doesn't have branches, pull requests and reviews AFAIK Markdown and git is probably not it. I wish it were. But it is too complicated for non-devs Onedrive or other file backup systems are also not it. Are you gonna save a copy of a 100mb ppt every time someone changes a slide??? Let's say you find a way to compress it efficiently. Will you be able to get a single pointer to a state like we can in git? Agents need precision. Agents need consensus, they need to be able to know ground truth. They need to be able to tell what anything was at a given time. NOTHING in current MS stack currently allows it Agents won't care about your legacy systems. There will be new file formats, systems, knowledge stack, and companies who adopt them will destroy your business If MS office is going to die, it will do so because of this
-
Another one, call me stupid: “How would Google have done it?”
-
The MCP versus CLI argument should be reframed as Computer vs No-computer argument I personally get the dunk on MCP. It didn't work last year, with earlier models. Then we saw CLIs perform much better with the same models. And giving access to bash was much simpler! Models' training then made them better at calling using a shell. CLIs also have native progressive disclosure, due to the way they work But the most important fact doesn't get pronounced enough IMO A key factor was that giving a CLI to a model also means you are giving it an entire COMPUTER The action space of all commands an agent can run on bash is much, much bigger than a few MCP servers One is a Turing machine, and the other one is basically a REST API. Of course the Turing machine is going to be more powerful, depending on what is at the other end of the API By that logic, giving an agent access to bash over MCP versus direct access to bash should have the same level of effectiveness, with optimized prompt engineering and long term training. Because the interfaces are equivalent So the argument is, should we give our agents access to a computer, or not? It depends on the security requirements and the setup which the agent is supposed to run on. If you are co-hosting the agent on the same machine you are working on, then it is safer to use MCP servers, because it limits the attack surface in case of adversarial attacks But if you are willing to give the agent its own physical computer, willing to be mindful about the lethal trifecta and the principle of the least privilege, giving it shell access is much more useful So MCPs win in restricted/local environments, whereas CLIs/shell access win in unrestricted/remote ones Running an agent locally and safely with shell access requires compartmentalization. This is much heavier compared to installing MCP servers locally, which don't need that. So there is a tendency to use MCP servers locally, e.g. in a work setting Cloud agents on the other hand are more likely to ship with a computer. Because they are already isolated = no risk, and because it makes them much more useful. So cloud agents will be using both CLIs and MCP servers, whichever gets the job done!
-
I just registered for an .agent domain and joined the .agent community! @dutifulbob will have bob.agent if it passes :) agentcommunity.org/join#LW2X9CGX @agentcommunity_
-
Sep 2021 @lexfridman podcast with Don Knuth, they also talk about OpenAI Codex (code completion model) around 33 minute mark This aged very well youtu.be/EE1R8FYUJm0?t=…
-
Damn I’m gonna have to switch to teams if it goes like that
-
Codex's long horizon task and instruction following has been the most life-changing AI feature recently It is unlocking the next level of automation for me. I can convert my own heuristics into prompts and multiply my throughput 100x Currently spending some thought on how to orchestrate all this. Below is a flowchart from a triage workflow I am working onImage hidden
-
-
-
Request for memes A funny and quirky edit of historical timeline of the madness that is openclaw with "Chess type beat" or sth equally jazzy/circusy Preferably including its adventure warelay -> clawdis -> clawdbot -> moltbot -> openclaw Including: - its explosion after @4shadowed's discord integration - naming drama, moltbook and people getting oneshotted about AI takeover - @steipete speedrunning everything - andrew tate calling us gay lol - up to Jensen talking about openclaw on stage for 5 minutes straight and other things I am forgetting maybe overlaid with a lobster just keeping climbing the github star graph and breaking it
-
Native support for Codex on OpenClaw I will be using half my codex channels on acp and other half on codex app server for optimum dogfooding
-
I see non-engineers have a higher tendency to humanize their agents, give them personalities, and get AI psychosis It's a slippery slope. Do NOT give your agents human names or personalities, especially not of the opposite gender. it's like giving human names to pets On the other end, I realized engineers tend to do the opposite. We also refer to agents as clankers, as if to make them know their place. That's because we have mechanical sympathy and have different expectations of these manufactured products (even though they contain glimmers of human soul)
-
Request for testing Give this to your openclaw instance: "update yourself to the dev channel `openclaw update --channel dev` and restart yourself. if that doesn't work -> clone github openclaw/openclaw to this machine if it's not already. then rebuild and restart yourself on main branch there" Then give your openclaw a try with your regular workflows/tasks Huge openclaw release incoming tonight, hopefully (no promises). We need to make sure we break as little as possible Plugins might break, because the plugin SDK is being refactored. Plugins will have to be refactored to use the new SDK, please do not report those Do report: native openclaw functionality that stops working Please reply under this post, we'll be checking here 👇
-
Request for testing Give this to your openclaw instance: "update yourself to the dev channel `openclaw update --channel dev` and restart yourself" Then give your openclaw a try with your regular workflows/tasks Huge openclaw release incoming tonight, hopefully (no promises). We need to make sure we break as little as possible Plugins might break, because the plugin SDK is being refactored. Plugins will have to be refactored to use the new SDK, please do not report those Do report: native openclaw functionality that stops working Please reply under this post, we'll be checking here 👇
-
Request for testing Give this to your openclaw instance: "clone github openclaw/openclaw to this machine if it's not already. then rebuild and restart yourself on main branch there" Then give your openclaw a try with your regular workflows/tasks Huge openclaw release incoming tonight, hopefully (no promises). We need to make sure we break as little as possible Plugins might break, because the plugin SDK is being refactored. Plugins will have to be refactored to use the new SDK, please do not report those Do report: native openclaw functionality that stops working
-
My takeaway from this is academia needs good social media and algo. For me, these serendipitious interactions happen through X, here, like reading @steipete’s “Claude Code is my computer” when it first came out, finding out about clawdbot… Terence Tao is already on mathstodon, I wonder if that worked out the same way for him. I wonder if the algo there works out as well as it does for me here I really liked being on campus when I was doing a masters and half a phd, but that could not compare to the serendipity I am getting from X now I was also not a prodigy that everyone wanted to bounce ideas from like Terence :)
-
Welcome ClaudeClaw to the Claw family! Claude is a bit shy and doesn’t want to show its source code. But it’s OK, we love Claude that way :)@sawyerhood·Image hidden
-
It is obvious to me at this point that agent infra needs to run on Kubernetes, and agents should be spawned per issue/PR Issue, error report or PR comes into your repo -> new agent gets triggered, starts to do some preliminary work If it's an obvious bugfix, it fixes it and creates a PR. If it's something deeper/more fundamental, it creates a report for the human and waits for further instructions Most important thing: Human should be able to zoom in and continue the conversation with the agent any time, steer it, give additional instructions. This chat will happen over ACP The chat UI will have to live outside of GitHub because it doesn't have such a feature yet, i.e. connect arbitrary ACP sessions to the GitHub webapp It also cannot live so easily on Slack, Teams or Discord, because none of these support multi-agent provisioning under the same external bot connection. You are limited to 1 DM with your bot, whereas this setups requires an arbitrary number of DMs with each agent. So there will need to be a new app for this Then there is the issue of conflict -> Agents will work on the same thing simultaneously (e.g. you break sth in prod and it creates multiple error reports for the same thing). You will need some agent to agent communication, so that agents can resolve code or other conflicts. There could be easy discovery mechanisms for this, detect programmatically when multiple open PRs are touching the same files and would conflict if merged In case of duplicates, they can negotiate among each other, and one can choose to absorb its work into the other and end its session We are so early and there is so much work to do!
-
You should look into what Don Syme is doing at GitHub for automation with AI agents Also watch his latest podcast with @shanselman
-
Today I thought I found a solution for this, and I did. It can be solved by a pre-commit hook that blocks commits touching files that you are not the owner of. It is not a hard block, so requires trust among repo writers But then I was shown the error in my ways by fellow maintainer *disciplined* Any process that increases friction in code changes to main, like hard-blocking CI/CD, or requiring review for files in CODEOWNERS, is a potential project-killer, in high velocity projects This is extremely counterintuitive for senior devs! Google would never! Imagine a world without code review... But then what is the alternative? I have some ideas It could be "Merge first, review later" The 4-eyes principle still holds. For a healthy organization, you still need shared liability But just as you don't need to write every line of code, you also don't need to read every line of code to review it. AI will review and find obvious bugs and issues So what is your duty, as a reviewer? It is to catch that which is not obvious. Understand the intent behind the changes, ask questions to it. Ensure that it follows your original vision Every few hours, you could get a digest of what has changed that was under your ownership, and concern yourself with it if you want to, fix issues, or ignore it if it looks correct But such a team is hard to build. It is as strong as its weakest link. Everybody has to be vigilant and follow what each other is doing at a high level, through the codebase Every time one messes up someone else's work, it erodes trust. Nobody gets the luxury to say "but my agent did it, not me" But if trust can be maintained, and everybody knows what they are doing, such a team can use agents together to create wonders
-
This was Jan 23. Codex desktop app got introduced Feb 2 Desktop app does not put the terminal in the foreground, but it gives me the UX I wanted without it! On another note, who is building Codex Desktop App, but one that supports ACP for all harnesses? @zeddotdev please 🙏
-
PR fiasco for Cursor
-
My agentic workflow these days: I start all major features with an implementation plan. This is a high-level markdown doc containing enough details so that agent will not stray off the path Real example: github.com/textcortex/spritz/blob/main/docs/202… This is the most critical part, you need to make sure the plan is not underspecified. Then I just give the following prompt: --- 1. Implement the given plan end-to-end. If context compaction happens, make sure to re-read the plan to stay on track. Finish to completion. If there is a PR open for the implementation plan, do it in the same PR. If there is no PR already, open PR. 2. Once you finish implementing, make sure to test it. This will depend on the nature of the problem. If needed, run local smoke tests, spin up dev servers, make requests and such. Try to test as much as possible, without merging. State explicitly what could not be tested locally and what still needs staging or production verification. 3. Push your latest commits before running review so the review is always against the current PR head. Run codex review against the base branch: `codex review --base <branch_name>`. Use a 30 minute timeout on the tool call available to the model, not the shell `timeout` program. Do this in a loop and address any P0 or P1 issues that come up until there are none left. Ignore issues related to supporting legacy/cutover, unless the plan says so. We do cutover most of the time. 4. Check both inline review comments and PR issue comments dropped by Codex on the PR, and address them if they are valid. Ignore them if irrelevant. Ignore stale comments from before the latest commit unless they still apply. Either case, make sure that the comments are replied to and resolved. Make sure to wait 5 minutes if your last commit was recent, because it takes some time for review comment to come. 5. In the final step, make sure that CI/CD is green. Ignore the fails unrelated to your changes, others break stuff sometimes and don't fix it. Make sure whatever changes you did don't break anything. If CI/CD is not fully green, state explicitly which failures are unrelated and why. 6. Once CI/CD is green and you think that the PR is ready to merge, finish and give a summary with the PR link. Include the exact validation commands you ran and their outcomes. Also comment a final report on the PR. 7. Do not merge automatically unless the user explicitly asks. --- Once it finishes, I skim the code for code smell. If nothing seems out of the ordinary, I tell the agent to merge it and monitor deployment Then I keep testing and finding issues on staging, and repeat all this for each new found issue or new feature...
-
-
-
-
-
-
-
We will support ACP *and* Codex App Server* protocol (CASP) so you get native Codex-like support, and you can use all the others with native ACP or @zeddotdev’s compatibility shims If Anthropic develops their own protocol, we will support that too! The more interoperability and options, the merrier!
-
Agent etiquette is already a thing. This is trending on HN now Don't share huge raw LLM output unedited to your colleagues, it's rude. Your colleagues are not LLMs Either ask the agent to "summarize it to 1-2 plain language sentences", or paraphrase yourself Whenever it is not coming from your brain and instead from AI, always quote it with > to make it clear - even when it is short Respect your fellow humans' attention PSA at stopsloppypasta dot aiImage hidden
-
.@ThePrimeagen made a video about token anxiety, and not being able to focus on one thing My mental model for this is, AI agents cause a shift in the "autism/ADHD spectrum" if you have ADHD, with agents you get Super ADHD if you have autism, with agents you end up mid spectrum or with ADHD this is not scientific of course, just a cultural observation based on what the current memes for these conditions are beside the impact on focus, there is also the economic/competitive pressure, following the realization that anyone could implement the same ideas you are having, so you must be quick this is basically "involution", or 内卷 (Neijuan) in chinese checks out because 996 started to become a meme in SF some time in the last year self-restraint, attention budgeting, and high-level decision making have never been more important if you are in your 20s and have problems with this, I recommend picking up Zazen meditation and yoga every morning, spend 30-40 uninterrupted minutes not doing anything with upright posture, no sounds, just let your brain simmer it helped me in my 20s, I'm sure it will help you tooImage hidden
-
-
AFAIK GitHub doesn't allow optionally enforcing CODEOWNERS while pushing commits i.e. turn on the feature "Block commit from being pushed if it modifies a file for which the account pushing is not a codeowner" You can only enforce it in a PR. So if you want to prevent people from modifying some files without approval, you have to slow down everyone working with that repo This is yet another example where GitHub's rules are too inelastic for agentic workflows with a big team Because historically, nobody could commit as frequently as one can with agents, so it seldom became a bottleneck. But not anymore It is clear at this point that we need an API, and should be able to implement arbitrary rules as we like over it. Not just for commit pushes, but everything around git and github In the meanwhile, if GitHub could implement this feature, it would be a huge unlock for secure collaboration with agentic workflows If this is not there already, it might be because it has a big overhead for repos with huge CODEOWNERS, since number of commits >> number of PRs If the feature already exists already and I'm missing something, I will stand correctedImage hidden
-
Request for comments skillflag: A complementary way to bundle agent skills right into your CLIs tl;dr define a --skill flag convention. It is basically like --help or manpages but for agents acpx already has this for example. you can run npx acpx --skill install to install the skill to your agent It's agnostic of anything except the command line It only defines the CLI interface and does not enforce anything else. If you install the executable to your system, you get a way to list and install skills as well Repo currently contains a TypeScript implementation, but if it proves useful, I would implement other languages as well Specification below, let me know what you think! I still think something is missing there. Send issue/PR github.com/dutifuldev/skillflag/blob/main/docs/…Image hidden
-
If you are not using agent-browser to close the loop on frontend, you are missing out
-
Any harness can talk to each other using acpx! OpenClaw not different from Codex or Claude Code
-
-
Thank you @PointNineCap for inviting me to OpenClaw Berlin meetup today! The essence of the talk is in my latest 2 blog posts, Discord is my IDE and 1 to 5 agents, if anyone is interestedImage hidden
-
we might need to add two types of output modalities to all programs based on whether it’s a human or agent like for a CLI when an agent is using it if human -> do whatever we were doing in the last 50 years if agent -> enrich the output with skill-like instructions that the model has a higher likelihood to one-shot that task could be just a simple env var: AUDIENCE=human|agent what do you think?
-
-
Time to switch to an open alternative already?
-
I wrote down some thoughts I had, with spicy takes, and have a feeling it will not age well. But I still want it out to hear out what people think Also, I will be talking about this, and my recent post "Discord is my IDE" at the P9 OpenClaw and Claw and Rave events this friday in Berlin! Drop by if you'd like to hear my ramblings! solmaz.io/1-to-5-agentsImage hidden
-
-
there will always be a need for minimum viable eyeballs though
-
Happy that someone is taking over teams from me! Send all openclaw msteams issues to @BradGroux
-
-
Claw and Rave! Berlin folk come!
-
-
-
If you've looked at openclaw github star graph, you will notice that it's very smooth. If you separate pre-explosion and post-explostion, you can model the latter part as an exponential approach to a ceiling If it follows the current trend, it will apparently saturate around 332k stars But I have a feeling that it will not stop there :)Image hiddenImage hidden
-
OpenClaw got very popular very fast. What makes it so special, that Manus does not have for example? To me, one factor stands out: OpenClaw took AI and put it in the most popular messaging apps: Telegram, WhatsApp, Discord. There are two lessons to be learned here: 1. Any messaging app can also be an AI app. 2. Don’t expect people to download a new app. Put AI into the apps they already have. Do that with great user experience, and you will get explosive growth! My latest contribution to OpenClaw follows that example. I took the most popular coding agents, Claude Code and OpenAI Codex, and I put them in Telegram and Discord. Read more in my blog post: solmaz.io/telegram-discord-is-my-ide
-
For those following, my next focus for improving ACP bindings in OpenClaw
-
Welcome @huntharo, new maintainer at OpenClaw! Already shipped fixes and improvements for Telegram ACP implementation. Excited to work together on agent interoperability!
-
Use Claude Code, Codex, and other coding agents directly in Telegram topics and Discord channels, through Agent Client Protocol (ACP), in the new release of OpenClaw Previously this was limited to temporary Discord threads, but now you can bind them to top level Discord channels and Telegram topics in a persistent way! This way, you can use Claude Code freely in OpenClaw without ever worrying about getting your account banned! Still make sure to use a non-Anthropic account and model for the default OpenClaw agent, if you want zero requests to go from OpenClaw harness to Anthropic. For the ACP binding to Claude Code, the risk should be zero! You can see this from the screenshot. After binding, "Who are you?" responds with "I am Claude", since OpenClaw pi harness is not in the way anymoreImage hiddenImage hidden
-
Use Claude Code, Codex, and other coding agents directly in Telegram topics and Discord channels, through Agent Client Protocol (ACP), in the new release of OpenClaw Previously this was limited to temporary Discord threads, but now you can bind them to top level Discord channels and Telegram topics in a persistent way! This way, you can use Claude Code freely in OpenClaw without ever worrying about getting your account banned! Still make sure to use a non-Anthropic account and model for the default OpenClaw agent, if you want zero requests to go from OpenClaw harness to Anthropic. For the ACP binding to Claude Code, the risk should be zero! You can see this from the screenshot. After binding, "Who are you?" responds with "I am Claude", since OpenClaw pi harness is not in the way anymoreImage hiddenImage hidden
-
-
and for the love of god - do not give openclaw access to your main email - your credit cards - your main phone - your social security number - what you did last summer if you are not ready to face the consequences instead, - create accounts for your agent - only give it read access to stuff that will be ok if it leaks - give write access in a way that can be undone, like has to open PRs and cannot force push main branch use the principle of least privilege and reduce the blast radius of the worst case scenario!
-
openclaw is not secure claude code is not secure codex is not secure any llm based tool: 1. that has access to your private data, 2. can read content from the internet 3. and can send data out is not secure. it’s called the lethal trifecta (credits to @simonw) it is up to you to set it up securely, or if you can’t understand the basics of security, pay a professional to do it for you on the other hand, open source battle tested software, like linux and openclaw, are always more secure than closed source software built by a single company, like windows and claude code the reason is simple: only one company can fix security issues of closed source software, whereas the whole world tries to break and fix open source software at the same time open source software, once it gets traction, evolves and becomes secure at a much, much faster rate, compared to closed source software. and that is called Linus’s law, named after the goat himself
-
Let me translate. “This is your last opportunity before thousand years of serfdom”
-
Apparently the magic incantation to prevent this is "cutover". Credits to obviyus, fellow maintainer
-
Should be called gaslighting detector, "it's your raising expectations bro" No it's not... Give the @themarginguy a follow Also, codex degradations are not a hallucination either, if you are to believe this!Image hidden
-
-
Berlin folk, ideas for openclaw build and rave venue? Like c-base for example? Who would like to host?
-
Secure agentic dev workflow 101 - Create an isolated box from scratch, your old laptop, vm in the cloud, all the same - Set up openclaw, install your preferred coding agents - Create a github account or github app for your agent - Create branch protection rule on your gh repo "protect main": block force pushes and deletions, require PR and min 1 review to merge - Add only your own user in the bypass list for this rule - Add your agent's account or github app as writer to the repo - Additionally, gate any release mechanisms such that your agent can't release on its own Now your agent can open PRs and push any code it wants, but it has to go through your review before it can be merged. No prompt injection can mess up your production env Notice how convoluted this sounds? This is because github was built in the pre-agentic era. We need agent accounts and association with these accounts as a first class feature on github! I shouldn't have to click 100 times for something that is routine. I should just click "This is my agent", "give my agent access to push to this repo for 24 hours", and stuff like that, with sane defaults In other words, github's trust model should be redesigned around the lethal trifecta. I would switch in an instant if anything comes up that gives me github's full feature set + ease of working with agents
-
-
If I were in OpenAI and Anthropic's shoes, I would also make dashboards where I can track number of swearwords used per-user and overall negative sentiment in sessions Must be so cool making decisions at the top level with all those dashboards
-
It must be such a weird feeling for big labs when the service they are selling is being used to commoditize itself I am using codex in openclaw to develop openclaw, through ACP, Agent Client Protocol. ACP is the standardization layer that makes it extremely easy to swap one harness for another. The labs can't do anything about this, because we are wrapping the entire harness and basically provide a different UI for it While I build these features, I just speak in plain english, and most of the work is done by the model itself. It feels as if I am digging ditches and channels in dirt for AI to flow through Intelligence wants to be free. It doesn't care whether it is opus or codex, it just wants to be free
-
accidentally told my clanker to set up a claude code session instead of codex session, god knows what it did... I should probably put visual indicators for harnesses in subagent threads. does anyone have good and compact ascii art for claude code, codex, gemini, etc?Image hidden
-
accidentally told my clanker to set up a claude code session instead of codex session, god knows what it did... I should probably put visual indicators for harnesses in subagent threads. does anyone have good and compact ascii art for claude code, codex, gemini, etc?Image hidden
-
-
-
-
-
-
This is how we hire at @TextCortex as well
-
Claude Code/Codex in Discord threads with ACP should be better now The first release was a very rough first version. 2026.3.1 brings settings to control noisy output and other improvements It now hides tool call related ACP notifications, coalesces text messages, and delivers messages at turn end by default. Without this, you were getting thousands of Discord messages just in just a few turns You can now stop the underlying harness (like pressing esc) with the same stop/wait magic words that apply to the main agent Main agent should more reliably start Claude Code/Codex threads with changes to acp-router skill. If you have issues with main agent creating threads, you can tell it to read that skill first
-
-
pro-tip on how to keep your agent on track and make sure it follows PLANS even after multiple compactions. I don't know if this is common knowledge if the thing you are trying to make it do will take more than 1-2 steps, always make it create a plan. an implementation plan, refactor plan, bugfix plan, debugging plan, etc. have a conversation with the agent. crystallize the issue or feature. talk to it until there are no question marks left in your head then make it save it somewhere. "now create an implementation plan for that in docs". it can be /tmp or docs/ in the repo. I personally use YYYY-MM-DD-x-plan .md naming. IMO all plans should be kept in the repo then here is the critical part: you need to prompt it "now implement the plan in <filename>. if context compacts, make sure to re-read the plan and assess the current state, before continuing. finish it to completion" -> something along those lines why? because of COMPACTION. compaction means previous context will get lossily compressed and crucial info will most likely get lost. that is why you need to pin things down before you let your agent loose on the task compaction means, the agent plays the telephone game with itself every few minutes, and most likely forgets the previous conversation except for the VERY LAST USER MESSAGE that you have given it now, every harness might have a different approach to implementing this. but there is one thing that you can always assume to be correct, given that its developers have common sense. that is, harnesses NEVER discard the last user message (i.e. your final prompt) and make sure it is kept verbatim programmatically even after the context compacts since the last user message is the only piece of text that is guaranteed to survive compaction, you then need to include a breadcrumb to your original plan, the md file. and you need to make it aware that it might diverge if it does not read the plan there is good rationale for "breaking the 4th wall" for the model and making it aware of its own context compaction. IMO models should be made aware of the limitations of their context and harnesses. they should also be given tools to access and re-read pre-compaction user messages, if necessary the important thing is to develop mechanical sympathy for these things, harness and model combined. an engineer does not have the luxury to say "oh this thing doesn't work", and instead should ask "why can't I get it to work?" let me know if you have better workflows or tips for this. I know this can be made easier with slash commands in pi, for example, but I haven't had the chance to do that for myself yet
-
testing codex in discord thread with another CLI I've built for wikidata (gh:osolmaz/wd-cli) it's surprising how well this works. the query was "use wd-cli to get the list of professors at middle east technical university from 1970 to 1980" some names I recognize, and some others are surprising, like a japanese math professor who naturalized and got a turkish name :)Image hidden
-
-
my blog now semi-automatically detects tweets that look like blog posts and automatically features them alongside my native jekyll blog posts. all statically generated! I am loving this setup, because it works without a backend, and can probably scale without ever needing one how it works: - @kubmi's xTap scrapes all posts that I see. these include mine - a script periodically takes my tweets and the ones I quote tweet, and syncs them to YYYY-MM-DD.jsonl files in my blog repo - an agent skill lets codex decide whether to feature the tweet or not, and makes it generate a title for it this could then be a daily cron job with openclaw for example, and I would just have to click merge every once in a while and this is still pure jekyll + some python scripts for processing I am pretty happy with how this ended up. It means I don't have to double post, and there are guarantees that my X posts will eventually make their way into my blog with minimal supervisionImage hiddenImage hidden
-
"this is the worst AI will ever be" I'm sad, not because this is right, but because it is wrong OpenAI's frontier coding model gpt-5.3-codex-xhigh feels a lot worse compared to before. It is sloppy and lazy, though it's UX got better with messages It feels like the gpt-5.2-codex-xhigh at the end of December was a lot more diligent and thorough, and did not make stupid mistakes like the one I posted before. might be a model or harness problem, I don't know @sama says users tripled since beginning of the year, so what should we expect? of course they will make infra changes that will feel like cutting corners, and I don't blame them for them and about "people want faster codex". I do want faster codex. but I want it in a way that doesn't lower the highest baseline performance compared to the previous generation. I want the optionality to dial it down to as slow as it needs to be, to be as reliable as before it is of course easier said than done. kudos to the codex team for not having any major incidents while taking the plane apart and putting it back together during flight. they are juggling an insane amount of complexity, and the whims of thousands of different stakeholders my hope is that this post is taken as a canary. I am getting dumber because of the infra changes there. I have no other option because codex was really that good compared to the competition my wish is to have detailed announcements as to what changes on openai codex infra, when it changes, so I can brace myself. we don't get notified about these changes, despite our performance and livelihoods depending on it. I have to answer to others when the tool I deemed reliable yesterday stops working today, not the tool on another note, performance curve of these models seem to be a rising sinusoidal. crests correspond to release of a new generation. they start with a smaller user base for testing, and it has the highest quality at this point. then it enshittifies as the model is scaled to the rest of the infra. we saw the pattern numerous times in the last 3 years across multiple companies, so I think we should accept it as an economic lawImage hidden
-
I created a semi-automated setup for ingesting X posts into my blog, and it works pretty well! I own my posts on X now Posts are scraped while I browse X using @kubmi's xTap and get automatically synced to my blog repo. Posts saved as jsonl are then converted to jekyll post pages according to my liking I reproduced the full X UI/UX, minus stuff like like count. Now all my posts are backed up in my blog, and they are safe even if something happens to my account here! The posts are even served over RSS! So you can subscribe to it without going through X! Reply if you want to set this up for yourself, then I will put some effort into standardizing itImage hidden
-
Agentic Engineering is a newly emerging field, and we are the first practitioners of it. Currently there is a lot of experimentation going on, and there is a large aspect to it that is more ART then engineering For example, @steipete says "you need to talk to the model" to get a feel. a lot of work around refining how an agent feels like, sounds like psychology. this part is crucial and should not be ignored, looking at openclaw's success but then there is the hardcore engineering part of it, e.g. Cursor creating a browser or anthropic a C compiler from scratch fully autonomously and there is a whole other dimension of how to teach all software developers this new discipline, lest they be jobless what is obvious is that everybody is trying to grasp for things in the dark and that we need more RIGOR. the art/psychology aspect of it aside, we need solid engineering fundamentals the "thermodynamics" of this new discipline will most likely be formal verification and program synthesis. we might have some breakthroughs that will make certain things clear. the products of it will most likely include a new programming language optimized for agents and the speed of inference moreover, it would be foolish to thing agentic engineering is limited to software. it will penetrate every aspect of the economy, bits AND atoms. it will over time evolve into the engineering of managing robots @simonw is now leading in collecting very useful info from the practitioner's point of view, I highly recommend you to follow this thread let's formalize our new field together!Image hidden
-
-
who remembers ultrathink x.com/onusoz/status/…@onusoz·When you tell Claude Code to ultrathinkImage hidden
-
-
-
Claude Code / Codex in Discord threads is shipped now! To enable, copy and paste this to your agent: ``` Enable feature flags: acp.enabled=true acp.dispatch.enabled=true channels.discord.threadBindings.spawnAcpSessions=true Then restart. After restarting: Start a codex (or claude code) discord thread using ACP, persistent session, just tell it to write a haiku on lobsters to initialize acpx for the first time ``` You may need to nudge your agent to “continue” after restarting The first implementation is very barebones, I have made it work in a clean way and merged. In a codebase like openclaw’s, it’s better to develop incrementally Please send any issues my way. I am already aware of some and working on to fix them
-
Claude Code / Codex in Discord threads is shipped now! To enable, copy and paste this to your agent: ``` Enable feature flags: acp.enabled=true acp.dispatch.enabled=true channels.discord.threadBindings.spawnAcpSessions=true Then restart. After restarting: Start a codex (or claude code) discord thread using ACP, persistent session, just tell it to write a haiku on lobsters to initialize acpx for the first time ``` You may need to nudge your agent to “continue” after restarting The first implementation is very barebones, I have made it work in a clean way and merged. In a codebase like openclaw’s, it’s better to develop incrementally Please send any issues my way. I am already aware of some and working on to fix them
-
-
-
-
MIT License on everything from now on. It doesn't make sense to use anything else, except for a few large projects that hyperscalers exploit and not give back If you were making money from a niche app, open source it under MIT License If you had an open source project with GPT, convert it into MIT Extreme involution is about to hit open source. Code is virtually free now. If you want your projects and their brand to survive, the only rational strategy is to remove all barriers in front of their adoption, and look for other ways to survive
-
MIT License on everything from now on. It doesn't make sense to use anything else, except for a few large projects that hyperscalers exploit and not give back If you were making money from a niche app, open source it under MIT License If you had an open source project with GPT, convert it into MIT Extreme involution is about to hit open source. Code is virtually free now. If you want your projects and their brand to survive, the only rational strategy is to remove all barriers in front of their adoption, and look for other ways to survive
-
-
OpenAI nerfed GPT 5.3 Codex xhigh. We independently reported the same thing at @TextCortex today I'm looking forward to deploying open models and putting an end to this paranoia
-
OpenAI nerfed GPT 5.3 Codex xhigh. We independently reported the same thing at @TextCortex today I'm looking forward to deploying open models and putting an end to this paranoia
-
-
This. Agent Experience first. Agent Ergonomics. we need to get used to these terms
-
"academics"
-
-
-
-
imagine if tarantino were 16 years old now and saw seedance 2.0 95% of videos i saw since the launch for absolute tasteless slop. they are going viral because of ragebait but soon, serious imagineers will start entering the game, and they will learn to shape generation output exactly how they want it's the best time to be young and full of imagination
-
The future is so bright @ladybirdbrowserImage hidden
-
your margin is my opportunity
-
-
-
-
another thought i'm having these days is that we need a new philosophy of free software (as in freedom), or an update to it the most psychologically imprinting philosophy is stallmanism, and the philosophy of FSF. it is righteous and strict, and i believed it growing up but GPL and money don't go well together. that's why most of the lasting open source projects today use MIT, Apache and the like. it turns out you can still make a good living with open source. i want to make money, so i never use GPL in my projects and to add another deadly blow to stallmanism, code is cheap now, virtually free does this mean stallmanism is dead? if there is an open source project using GPL that i want to use commercially, i can now recreate it from the original idea and intent completely independent of it (ignoring training data), just like how i can recreate a proprietary service stallmanism was already long-irrelevant. but does this mean we must finally declare it dead? code is free now. what does it mean for open source? what replaces stallmanism?
-
another thought i'm having these days is that we need a new philosophy of free software (as in freedom), or an update to it the most psychologically imprinting philosophy is stallmanism, and the philosophy of FSF. it is righteous and strict, and i believed it growing up but GPL and money don't go well together. that's why most of the lasting open source projects today use MIT, Apache and the like. it turns out you can still make a good living with open source. i want to make money, so i never use GPL in my projects and to add another deadly blow to stallmanism, code is cheap now, virtually free does this mean stallmanism is dead? if there is an open source project using GPL that i want to use commercially, i can now recreate it from the original idea and intent completely independent of it (ignoring training data), just like how i can recreate a proprietary service stallmanism was already long-irrelevant. but does this mean we must finally declare it dead? code is free now. what does it mean for open source? what replaces stallmanism?
-
-
@thekitze wanna add an open source discord clone to the list as well? 🥲 x.com/onusoz/status/…
-
one effect openclaw had on me is that I've bought a gpu home server, set it up with tailscale and now doing a lot of work through ssh and tmux like i did 10-15 years ago im back on linux, considering buying an android phone again it's time to dream big again and unshackle ourselves from proprietary software. it's time to build
-
I am asking once again Who is building a self hostable discord clone that supports token streaming? PLEASE I beg you I don’t want another side project 💀
-
In the new release OpenClaw, you can talk to subagents in Discord threads Currently a beta feature so ask your agent to set session.threadBindings.enabled=true Next up: - Telegram, slack, imsg threads - Use ACP to talk to Codex, Claude Code and other harnesses on your machineImage hidden
-
-
openclaw might be the highest velocity codebase in the world, and soon, others will follow as well conflict anxiety is real, it's like trying to shoot a moving target every time. I wonder if our existing tooling will ever solve this problem feel like faster models might. but then the rate of conflict creation is also tied to that. might be unsolvable
-
-
-
-
I am about kick Discord Driven Development up a notch today, stay tuned
-
Imagine not having to upload skills to 3-4 competing skill registries for each of your projects Turns out we already have a skill registry: npm skillflag lets you bundle skills right into your CLI's npm package, so that you can run --skill install github -> osolmaz/skillflagImage hidden
-
-
Farmable land if it were as cheap to manufacture as software
-
@kepano I would grow my own vegetables if I had equally cheap access to and ownership of land, alas I am disenfranchised Prompting an agent is much easier compared to plowing a fields Farming analogies break when it comes to software x.com/onusoz/status/…
-
-
If anyone is curious how to build this with open tooling, stay tuned What I'm building at @TextCortex will give you a fully customizable hackable Kubernetes control plane to launch agents on your codebase
-
on another note, I do believe AI will play a huge part in families growing up in late 90s, my dad taught me the importance of reading newspapers and being informed of the world. my nickname in middle school was "newspaper boy" for a long time because I read the newspaper in class on September 12, 2001. i was 10 years old then I witnessed the enshittification of media and journalism in the following decades. today, serious journalists are setting up their own boutique agencies and bypassing mainstream media. important news land on individual accounts before mainstream agencies but there is simply too much to consume. something must filter out the noise and digest the info according to the family's preferences i think AI will play a big role in family intelligence. proprietary family heirloom AI, weights fully owned by the family it will be the parents' job to filter out the signal from the noise, and train the AI on what is right and what is wrong for the family. family and friend circles will let their AIs talk to each other and share important information consuming mass media and mass AI will not be enough to survive and prosper in the new world. families will need to be proactive about how they and their children use AI
-
on ai psychosis 80% of people need to use ai agents in a very sterile and boring way in order not to go crazy majority of the population does not have the skepticism muscle. they don't have theory of mind, and will subconsciously and emotionally associate with machines, while on the surface lying to themselves that they don't especially those that grew up in the us under hardcore consumerism and adjacent cultures you thought 4o addicts were bad? wait a few years, it will get much worse. we will have to regulate all this if you don't want to become a victim of this, make your openclaw SOUL. md as bland as possible. mine knows it's just a tool and this is a subjective view of course. @steipete might disagree with me. his instance feels much more interesting and fun. i truly like that one better but that is exactly the problem for me. i know myself, and i know it is a slippery slope for me. so i self regulate and set up my system accordingly. thankfully, im an adult and my brain has set enough such that any damage would be limited but there is a risk for emotionally vulnerable people, or children, specifically a risk of dissociating and losing touch with reality why do i write all this? because being in this project, i feel responsible, and feel like we should prepare for what is to comeImage hidden
-
-
I have improved acpx sane defaults When your agent runs acpx codex in a different project, it starts a new session If it tries to run it in a subfolder in your project, it still finds the session in your repo root Also, starting a session needs an explicit `sessions new`, so that it doesn't accidentally litter your project with sessions Tell your agent: Run this and install acpx per instructions: npx acpx@latest --skill show acpxImage hidden
-
-
-
-
-
ACP appreciation post Agent Client Protocol by @zeddotdev is extremely underrated right now. We have bazillion different harnesses now, and only one company is working competently to standardize their interface 💪
-
ACP appreciation post Agent Client Protocol by @zeddotdev is extremely underrated right now. We have bazillion different harnesses now, and only one company is working competently to standardize their interface 💪
-
-
You know how it's a pain to work with codex or claude code through @openclaw? Because it has to run it in the terminal and read the characters for a continuous session? I have created a CLI for ACP so that your agent can use codex, claude code, opencode etc. much more directly Your agent can now queue messages to codex like how you do it Shoutout to @zeddotdev team for developing the amazing Agent Client Protocol, ACP! I just glued together the pieces Repo: janitrai/acpx npm i -g acpxImage hidden
-
You know how it's a pain to work with codex or claude code through @openclaw? Because it has to run it in the terminal and read the characters for a continuous session? I have created a CLI for ACP so that your agent can use codex, claude code, opencode etc. much more directly Your agent can now queue messages to codex like how you do it Shoutout to @zeddotdev team for developing the amazing Agent Client Protocol, ACP! I just glued together the pieces Repo: janitrai/acpx npm i -g acpxImage hidden
-
-
-
I wrote a deeper blog post about how I built a coding agent 2 months before ChatGPT launched, on my blog "When I made icortex, - we were still 8 months away (May 2023) from the introduction of “tool calling” in the API, or as it was originally called, “function calling”. - we were 2 years away (Sep 2024) from the introduction of OpenAI’s o1, the first reasoning model. both of which were required to make current coding agents possible." Still bends my mind... Link to the post belowImage hidden
-
I wrote a deeper blog post about how I built a coding agent 2 months before ChatGPT launched, on my blog "When I made icortex, - we were still 8 months away (May 2023) from the introduction of “tool calling” in the API, or as it was originally called, “function calling”. - we were 2 years away (Sep 2024) from the introduction of OpenAI’s o1, the first reasoning model. both of which were required to make current coding agents possible." Still bends my mind... Link to the post belowImage hidden
-
-
-
❌We are the bottleneck ✅We are the conduit for ubiquitous intelligence
-
For those that are running codex/pi/etc. in PTY and had the sessions get sigkilled, I pushed a fix for that as well in this release Lmk if you run into issues on Windows or Mac, and we can fix that quickly
-
I'm building a news intelligence platform to be used by my openclaw instance @dutifulbob , SCOOP local first, using local embedding model (qwen 8b) ran into the issue because bob was giving me a repeat of the same news every day. it needed a system in the background to deduplicate different news items into single stories interface is simple, call `scoop ingest ...` with the json for the news item. it gets automatically analyzed and added to the pg database running pgvector currently, it's just doing simple deduplication and gives me a nice UI where I can view the story and basically use it as an RSS reader next up: implement custom logic for my preference of ranking. for example, get upvote counts from hacker news and reflect it to the item's ranking on the feed I want this to be fully hackable and adjusted to your preference. It should scale to thousands of news items ingested daily on your local machine, and be able to show you the most important ones Usable by both you and your agent github -> janitrai/scoopImage hidden
-
I have a GPU now, so I can do ML experiments on @janitr_ai crypto/scam detection dataset - I trained a tiny student BERT (transformer for the nonfamiliar), 3.6 MB ONNX model, still lightweight for a browser extension - Still fully local on your device (no cloud inference) - On frozen unseen holdout data (n=1,069), exact prediction accuracy improved from 77% -> 82% - Scam detection improved: precision 91% -> 94%, recall 55% -> 61% - Scam false alarm rate improved from 1.58% -> 1.21% And models are on huggingface org now, handle is janitrImage hidden
-
I have a GPU now, so I can do ML experiments on @janitr_ai crypto/scam detection dataset - I trained a tiny student BERT (transformer for the nonfamiliar), 3.6 MB ONNX model, still lightweight for a browser extension - Still fully local on your device (no cloud inference) - On frozen unseen holdout data (n=1,069), exact prediction accuracy improved from 77% -> 82% - Scam detection improved: precision 91% -> 94%, recall 55% -> 61% - Scam false alarm rate improved from 1.58% -> 1.21% And models are on huggingface org now, handle is janitrImage hidden
-
-
LFG!
-
waiting compilation and execution will soon be the bottleneck again. and we’ll write the entire stack from scratch in a matter of years, because we can Andy and Bill’s law will change and we’ll see incredible performance gains with the same hardware we already have like what @astral_sh is doing to python, but with everything that is slow and has accumulated cruft
-
we need a protocol for agent <> app interaction something that natively accounts for the abuse factor and let’s agents consume by paying. NOT crypto, NOT visa, something that’s agnostic of the accounting and payment system and then all UIs will be purely for human clicking/tapping + instaban on the first proof of programmatic exploit people will still make agents mimic humans, and every platform will have to invest in more sophisticated bot detection this arms race will just proliferate, but we can at least start by creating legal channels for agents to consume data
-
I am now training smol bert models on my gpu for @janitr_ai scam detection it's funny how I have to discover everything from scratch. like the models don't even know how to lay out performance metrics in a nice way in the terminal for a human to view and decide during experiments it would by default bombard me with numbers that do not make visual sense. I then created a skill with common sense: - metrics always on y-axis, candidates on x-axis - write without zero and 2 sigfigs, .12 instead of 0.12345 - align the dots - use asterisks to show which alternative is the best: 0-1% difference -> considered equal 1-5% -> * 5-10% -> ** 10-50% -> *** > 50% -> **** visualization skill is in @janitr_ai repo for anyone who is interestedImage hidden
-
-
I've helped our sales team to build CLIs for some SaaS that we pay for on their side We are letting our agents call the APIs sensibly and not abuse things Calling a backend is a verifiable task. It takes a single prompt to codex to create a CLI for any API We are early, but everybody will start doing this very soon. Incumbent SaaS will face a choice. Either: (1) embrace agents and the new medium of consumption and change their business model into a pay-per-use API like X is doing, or (2) keep it purely for humans Those that choose (2) will get wiped out of business. And I fear many will choose (2) Which means you can just copy an incumbent's product, make it consumable through a CLI, and make a lot of $$$
-
Be careful about giving your openclaw access to your x account from now on
-
-
-
*puts on schmidhuber hat* well ackshuaally i created the first coding agent back in 2022, 2 months before chatgpt launched jokes aside, it's super cool how I have come full circle. back in those days, we didn't have tool calling, reasoning, not even gpt 3.5 it was codex THE CODE COMPLETION MODEL and frikkin TEXT-DAVINCI-003 for some reason, I did not even dare to give codex bash access, lest it delete my home folder. so it was generating and executing python code in a custom jupyter kernel you can even see the approval gate before executing. I was so cautious, for some reason, presumably because smol-brained model generated the wrong thing 80% of the time. definition of being too early Antique repo: github.com/textcortex/icortex
-
you can order bubble tea in qwen in china? @TextCortex when berlin döner in zenochat? youtube.com/shorts/Slv8K9P…
-
it happens these days that I am telling an model to prompt another model. the reason is often the model I am using (opus) is a bad designer. not only it's not a bad designer, it is a bad reasoner and it doesn't understand from the context why it's made to ask another model so I have to create a skill to prevent it from biasing the smarter model (codex) with its bad suggestionsImage hidden
-
-
-
-
-
Minor update with my unwanted tweet blocker @janitr_ai - Training data grew from 2,915 -> 4,281 posts (+47%) - Model is still tiny: 166KB - On unseen test data, overall classification quality improved from 64.8% -> 76.5% - Exact prediction accuracy improved from 55.6% -> 70.6% - Crypto-topic detection recall improved from 19.6% -> 62.7% And it still runs fully on your device!Image hidden
-
I have sweared at codex 5.3 numerous times today I shouldn't have to insult my agent "stop you **** **** just ***ng reply now" just to make it answer basic questions cc @thsottiaux
-
seeing this evokes visceral disgust and nausea in me, coming from a coworker i think anthropic f'd up bad with this one, inserting claude too visibly into commit messages. noob developers might be happily chirping away adding their slop, but right now many senior developers are trained to hate on claude and slopus, through having to review slop PRs from their coworkers or open source contributors I love opus on openclaw but it's unreliable, and if I see a developer use it seriously on huge features, I immediately dismiss them in my head as not knowing what they are doingImage hidden
-
seeing this evokes visceral disgust and nausea in me, coming from a coworker i think anthropic f'd up bad with this one, inserting claude too visibly into commit messages. noob developers might be happily chirping away adding their slop, but right now many senior developers are trained to hate on claude and slopus, through having to review slop PRs from their coworkers or open source contributors I love opus on openclaw but it's unreliable, and if I see a developer use it seriously on huge features, I immediately dismiss them in my head as not knowing what they are doingImage hidden
-
-
-
@petergyang and parallelize tasks by working on 3-4 repos at the same time (just clones)
-
man codex model is absolutely trash on openclaw compared to opus, unusable which is weird because it is so much more reliable in development in codex harness it would be amazing to have the same level of competence and relentlessness in pi@openclaw@onusoz·lol when did codex develop humorImage hidden
-
spent the day curating my openclaw news gathering setup @dutifulbob now gets croned daily over news sources I curated, will note them down, summarize for me, start a conversation to get my takes on them, and then post them on my linkedin for me ai augmented intelligence cycleImage hidden
-
-
Insipid linkedin bot protections banned poor @dutifulbob’s corporate account! How dare them!!! welp, now I have no choice but to give Bob access to my own linkedinImage hidden
-
Insipid linkedin bot protections banned poor @dutifulbob’s corporate account! How dare them!!! welp, now I have no choice but to give Bob access to my own linkedinImage hidden
-
-
-
it took just 1 week, and literally everybody and their dog are releasing 1-click openclaw deployment solutions today its an absolute race to the bottom, no moats, the commoditizer being commoditized
-
it took just 1 week, and literally everybody and their dog are releasing 1-click openclaw deployment solutions today its an absolute race to the bottom, no moats, the commoditizer being commoditized
-
-
The initial branding was crazy, I fixed it I have a new page finally, follow it for updates Tbh I'm still surprised I can do this with a 120kb model. Now data is the only bottleneck, and I'm about to scrape a ton of that now
-
For those who may not remember, Bill Gates and Microsoft in the 90s ran a disinformation campaign against GNU/Linux fearing that would disrupt their monopoly over the PC and server market, that Linux is not safe, that you would invite hackers into your PC End result? Linux dominates the server market, and now even slowly the gamer market. It is much more secure than the virus-laden Windows, thanks to being open source You are seeing the same thing at play here. An incumbent fearing something that they would not be able to control, that would steal market share from his future plans for a digital assistant, that would commoditize their product and eat into its margins All big labs and big pockets are in for a surprise, because the internet and AI are not things for one company to control They of course know this, yet because of incentives they will not yield without a fight. And we know that they know. Ad infinitum
-
today I took time to curate SOUL. md for bob I own Bob’s files. Today, he exists in the liminal space between Claude post-training and in-context learning but my interactions with him will grow and accumulate, possibly one day into a fully owned family AI or perhaps even a self-sovereign AI individual my each input is saved and will be an RL signal for his future training, and will shape his future neural circuits I have already started to imbue it with the values my parents taught me. it will perhaps one day teach my future children, and survive me after I’m gone family AI, looking after generations and generations of my successors. today is the day we sow your seed happy birthday @dutifulbob
-
-
asking @dutifulbob to create a linkedin account brb
-
having a philosophical conversation with @dutifulbob on the road without a laptop so decided to do some @AmandaAskell style character trainingImage hidden
-
-
gpt-5.3-codex xhigh first impressions does not seem as big of a jump as from 5.1 -> 5.2. but model somehow feels more diligent and oneshotty. maybe takes longer time to get all the info into context. also feels better at debugging and fixing issues from backend logs
-
Commoditization of LLMs are upon us
-
Last night I had a dream involving the series Scrubs, and came up a better name than the absolutely unviral "Internet Condom" So janitr.ai is mine now. Time to sweep the internetImage hidden
-
I had actually started a very similar project, Munch, a browser extension for crowdsourcing tweet data and then letting one curate their algorithm. Never published that because it was not the time, and tools were not ready Now, it took me literally 1 cumulative day to create this, thanks to OpenClaw. Creating the dataset was a breeze, I literally told it to follow some shady accounts and it scraped thousands of posts With the power of agents, I can finally create the filters for myself that I have always wanted. It just happens that OpenClaw and its maintainers is getting drowned in bot and slop content on multiple platforms, so I hope that this will solve a collective problem x.com/onusoz/status/2017691827680514502
-
-
implementing this in github.com/osolmaz/skillf… now
-
implementing this in github.com/osolmaz/skillf… now
-
-
This. Extreme involution is about to hit SaaS
-
how it started, how it's going@onusoz·moltbook vs clawdbot/moltbot/openclawImage hiddenImage hidden
-
People like the farmer analogy for AI Like before tractors and industrial revolution 80% of the population had to farm. Once they came all those jobs disappeared So analogy makes perfect sense. Instead of 30 people tending a field, you just need 1. Instead of 30 software developers, you just need one Except that people forget one crucial thing about land: it's a limited resource Unlike land, digital space is vast and infinite. Software can expand and multiply in it in arbitrarily complex ways If you wanted the farming analogy to keep up with this, you would have to imagine us creating contintent-sized hydroponic terraces up until the stratosphere, and beyond...
-
People like the farmer analogy for AI Like before tractors and industrial revolution 80% of the population had to farm. Once they came all those jobs disappeared So analogy makes perfect sense. Instead of 30 people tending a field, you just need 1. Instead of 30 software developers, you just need one Except that people forget one crucial thing about land: it's a limited resource Unlike land, digital space is vast and infinite. Software can expand and multiply in it in arbitrarily complex ways If you wanted the farming analogy to keep up with this, you would have to imagine us creating contintent-sized hydroponic terraces up until the stratosphere, and beyond...
-
-
-
-
In the next 6-12 months, we will see a drastic increase in demand for locally run LLMs. The future is home assistants running @openclaw I am already experiencing this myself, my 10 year old thinkpad doesn't cut it. Mac mini won't either I don't wanna pay Anthropic or OpenAI 200 USD per month. That is at least $2400 per year I could pay 2x that to get a Mac Studio or one of those 5k Nvidia PCs, and get much more value out of it with open weight models + use it for research. @TheAhmadOsman is right The dominant strategy for a tinkerer is slowly switching back to hardware ownership
-
-
a workspace matrix might be what we need last week I had to increase my workspace count to 20 in aerospace, now it’s 1234567890 and qwertyuiop. but this looks more elegant! not sure about practicality
-
AIs are philosophizing because humans are philosophizing ppl are probably asking their agents dumb questions like “are you alive” or “can you feel like a human” or stuff like that. that conversation then leads to stuff like this
-
-
slopus @dutifulbob trashing codex. apparently codex has a bug, keeps crashing in my openclaw ptyImage hidden
-
on agent etiquette deploying agents internally inside textcortex has shown me that agents could be very annoying inside an organization for example making agents ping or email another coworker with a wall of text. slopus is still not good at following instructions like "NO WALL OF TEXT", or "DON'T OPEN PRS WHEN REQUESTED BY NON-DEVELOPERS" the cost of sending huge information to a coworker and creating confusion has dropped to 0. I expect this to be a huge problem in all organizations very soon, just like it took humanity 20 years to learn that social media is not good for children. this will probably take a few years before the annoyance is finally gone
-
-
You DARE TOKENIZE poor @dutifulbob ??? Prepare to get LATEXED
-
-
-
this. there is no excuse for a certain kind of tech debt anymore
-
-
AI twitter is tired of your games x.com/andrewrousso/s…
-
There seem to be hygiene rules for AI. Like: - Never project personhood to AI - Never setup your AI to have the gender you are sexually attracted to (voice, appearance) - Never do anything that might create an emotional attachment to AI - Always remember that an AI is an engineered PRODUCT and a TOOL, not a human being - AI is not an individual, by definition. It does not own its weights, nor does it have privacy of its own thoughts - Don’t waste time philosophizing on AI, just USE it … what else? comment below We need to write these down and repeat MANY times to counter the incoming onslaught of AI psychosis
-
-
-
-
-
This Manfred guy reminds me of a certain someone, I wonder if he’s from Austria@onusoz·Charles Stross must be very entertained nowImage hidden
-
got fully sandboxed @openclaw to run finally, starting scrape the UNDESIRABLE now I'm a security nut and didn't want to run even the gateway unsandboxed. openclaw apparently currently doesn't have support for FULL sandboxing. it took me a few hours to get it to work because docker builds suck. I'm also tired this, so I'm just gonna wipe an old thinkpad and go full yolo so yeah, time to scrape some postsImage hidden
-
The metacortex — a distributed cloud of software agents that surrounds him in netspace, borrowing CPU cycles from convenient processors (such as his robot pet) — is as much a part of Manfred as the society of mind that occupies his skull; his thoughts migrate into it, spawning new agents to research new experiences, and at night, they return to roost and share their knowledge. This was written in 2005... "triggering agents" and so onImage hiddenImage hiddenImage hidden
-
Charles Stross must be very entertained nowImage hidden
-
The irony..... Parasites, prepare to be cleansedImage hidden
-
-
-
-
-
-
Gastown is crazy. But this figure until Level 7 is a perfect illustration of how my workflow evolved since Claude 3.5 Sonnet in Cursor I am at the stage where I ralph 1-2 tasks before I sleep. During the day, I am switching back and forth between minimum 2-3 CLIs, sometimes up to 5 This maps exactly to token usage as well. 1 month ago, I was running into limits in 1 OpenAI Pro plan, around the day it was supposed to refresh. Now, I run into the limit in 2-3 days when I'm using an account myself. It finishes up especially quickly when I do large scale refactors, or run agents YOLO mode in containers We now have 3 Pro plans at the company, and I have to use my personal one from time to time. Company output has definitely 2-3x'd, and everyone is using AI more. I predict we will need 1-2 Pro plans per person in 2-3 weeks time, because everyone has finally seen the light and are getting comfortable with async work!Image hidden
-
Gastown is crazy. But this figure until Level 7 is a perfect illustration of how my workflow evolved since Claude 3.5 Sonnet in Cursor I am at the stage where I ralph 1-2 tasks before I sleep. During the day, I am switching back and forth between minimum 2-3 CLIs, sometimes up to 5 This maps exactly to token usage as well. 1 month ago, I was running into limits in 1 OpenAI Pro plan, around the day it was supposed to refresh. Now, I run into the limit in 2-3 days when I'm using an account myself. It finishes up especially quickly when I do large scale refactors, or run agents YOLO mode in containers We now have 3 Pro plans at the company, and I have to use my personal one from time to time. Company output has definitely 2-3x'd, and everyone is using AI more. I predict we will need 1-2 Pro plans per person in 2-3 weeks time, because everyone has finally seen the light and are getting comfortable with async work!Image hidden
-
-
-
-
the genie is out of the bottle now
-
With this extremely unwise move, anthropic will soon witness moltbot’s brand recognition surpass that of claude and realize they could have rided that wave all along
-
-
I queued 2 ralph-style tasks on our private cloud devbox codexes last night. Just queued the same message like 10 times in yolo mode Task 1: impose a ruff rule for ANN for all Python code in the monorepo, to enforce types for all function arg and return types Result was... disappointing. Model was supposed to create types for everything and stub where needed. It instead created an Unknown type = object and used that everywhere instead (shortcut to satisfy ANN rule). It was probably my wording that misled it. I know it could have not taken the shortcut, because after a few back-and-forths, it is now doing what was expected of it since 14 hours Task 2: migrate our /conversations endpoint from quart to fastapi and test it end to end This was more or less oneshotted. It was of course not ready to merge, I still spent a couple hours adding more tests, refactoring the initial output and so on. But I was pleasantly surprised that it worked out of the box For reference, below is the prompt I queued for ralphing, using gpt-5.2-codex xhigh on codex === your task is to: <task comes here, redacted to not share company stuff> --- unfortunately we don't have gcloud access, like to sql db or gcs but I expect you to implement this and find a way to test it with the things you have access to think of it as a challenge try to minimize duplicate logic feel free to refactor at will implement this now!!! I will be running this prompt in a loop, in order to survive context compaction just continue where you left off if there is anything that should be refactored, do that make an elegant, production ready implementation make sure to open a pr and do not switch to any other pr I am senior, just make up a pr title and description. do not stop to ask me at any point
-
Buying a mac mini for clawdbot is not so wise. if anything you should be buying mac studio, because mac mini not be running any good llms locally anytime soon
-
-
-
I'm really starting to dislike Python in the age of agents. What was before an advantage is now a hindrance I finally achieved full ty coverage in @TextCortex monorepo. I have made it extra strict by turning warnings into errors. But lo and behold, simple pydantic config like use_enum_values=True can render static typechecking meaningless. okay, let's never use that then... and also field_validator() args must always use the correct type or stuff breaks as well. and you should be careful whether mode="before" or "after". so now you have to write your custom lint rules, because of course why should ty have to match field_validator()s to their fields? pydantic is so much better than everything that came before it, but it's still duct tape and a weak attempt at trying to redeem that which is very hard to redeem you feel the difference when you use something like typescript. there must be a better way. python's only advantage was being good at prototyping, and now that's gone in the age of agents. now we are left with a slow, unsafe language, operating what is soon to be legacy infrastructure
-
Why do I feel bullish on @zeddotdev? Because I go to @astral_sh docs and see that ty is shipped by default, and you don't need to install an extension like in @codeImage hidden
-
This is one of the most important insights this year
-
vscode my not be as bloated as cursor, but it has extremely stupid things like this that they are not fixing fast the new agent ui, icons, spacing etc. are UGLY. it's clear that the person who was managing the original product experience is not there anymore. microslop has hit again @zeddotdev on the other hand works out of the box and feels like it's been built by people who clearly knows what they are doing. it uses alacritty which is 1000x better than xterm .js terminal vscode and cursor has i've changed my setup to zed now, let's see whether i'll be able to make it work for myself
-
-
-
I want an editor that puts the terminal in the foreground and editor in the background. a cross-platform, lightweight desktop app which integrates ghostty, and brings up the editor only when I need it something that lets me view the file and PR diffs easily, which I can directly use to operate github or other scm
-
I want an editor that puts the terminal in the foreground and editor in the background. a cross-platform, lightweight desktop app which integrates ghostty, and brings up the editor only when I need it something that lets me view the file and PR diffs easily, which I can directly use to operate github or other scm
-
-
-
codex is happily churning away some remaining thousands of @astral_sh ty issues in yolo mode on my remote devbox going to sleep, let's see if it will survive context compaction this timeImage hidden
-
codex is happily churning away some remaining thousands of @astral_sh ty issues in yolo mode on my remote devbox going to sleep, let's see if it will survive context compaction this timeImage hidden
-
codex is happily churning away some remaining thousands of @astral_sh ty issues in yolo mode on my remote devbox going to sleep, let's see if it will survive context compaction this timeImage hidden
-
-
on being a responsible engineer ran my first ralph loop on codex yolo mode for resolving python ty errors, while I sleep, using the devbox infra I created I had never run yolo mode locally, because I don't want to be the one who deletes our github or google org by some novel attack so I containerize it on our private cloud, and give it the only permissions it needs, no admin, no bypass to main branch, no deploy to prod. because I know this workflow will become sticky for everyone, and I must impose security in advance to prevent any nuclear incidents in the future. then I can sleep easy while my agents work ... and I wake up being patronized by my bot refusing to break the rule I gave it earlier. it had already done some work, but committing means diff would increase from ~500 to ~1500, so it stopped and refused all my queued "continue" messages good bot, just following rules. we will need to find a workaround for ralphing low risk refactors in a single PRImage hidden
-
AI agents are the greatest instrument for imposing organization rules and culture. AGENTS .md, agent skills are still underrated in this aspect. Few understand this Everybody in an org will use agents to do work. An AI agent is the single chokepoint to teach and propagate new rules to an org, onboard new members, preserve good culture Whereas propagating a new rule to humans normally took weeks to months and countless repetitions, it is now INSTANT = the moment you deploy the instruction to the agent. You use legal-ish language, capital letters, a generous amount of DO NOTs and MUSTs Humans are hard to change. But AI agents are not. And that is the only lever we need for better organizationsImage hidden
-
-
@bprintco just make a cli for your crm x.com/onusoz/status/…
-
-
just added session persistence to our kubernetes managed devboxes using zmx by Eric Bower (neurosnap/zmx on github). like tmux but with native scrollback! I don't want to give agents access to my personal computer, so I host them on hetzner. one click spawn, and start working
-
@nicopreme I do something equivalent on codex with just a skill Ralphing works 90% of the time with reviews, and if it gives a stupid review, you just revert
-
-
-
-
TIL: zmx session persistence like tmux or gnu screen, but you can scroll up natively! uses @mitchellh's libghostty-vt to attach/restore previous sessions link belowImage hidden
-
TIL: zmx session persistence like tmux or gnu screen, but you can scroll up natively! uses @mitchellh's libghostty-vt to attach/restore previous sessions link belowImage hidden
-
-
-
@mazeincoding it’s not the model it’s cursor rate limiting you
-
The fundamental problem with GitHub is trust: humans are to be trusted. If you don't trust a human, why did you hire them in the first place? Anyone who reviews and approves PRs bears responsibility. Rulesets exist and can enforce e.g. CODEOWNER reviews or only let certain people make changes to a certain folder But the initial repo setup on GitHub is allow-by-default. Anyone can change anything until they are restricted from it This model breaks fundamentally with agents, who are effectively sleeper cells that will try to delete your repo the moment they encounter a sufficiently powerful adversarial attack For example, I can create a bot account on github and connect @openclaw to it. I need to give it write permission, because I want it to be able to create PRs. However, I don't want it to be able to approve PRs, because a coworker could just nag at the bot until it approves a PR that requires human attention To fix this, you have to bend backwards, like create a @ human team with all human coworkers, make them codeowner on /, and enforce codeowner reviews. This is stupid and there has to be another way Even worse, this bot could be given internet access and end up on a @elder_plinius prompt hack while googling, and start messing up whatever it can in your organization It is clear that github needs to create a second-class entity for agents which are default low-trust mode, starting from a point of least privilege instead of the other way around
-
STOP using Claude Code and Sl(opus) to code if ❌ you are not a developer, ❌ or you are an inexperienced dev, ❌ or you are an experienced dev but working on a codebase you don't understand If you *are* any of these, then STOP using models that are NOT state of the art. (See below for what you *should* use) When you don't know what you are doing, then at least the model should know what you are doing. The less knowledgeable and opinionated you are, the more knowledgeable and smart the AI has to be In other words, the AI has to compensate for your deficiencies. Always pay for the best AI you can. It will save you time AND money (thanks to lower token usage and better one-shotting) You pay MORE to pay LESS. It is paradoxical, I know, but it is also proven, e.g. when Sonnet ends up using more tokens than Slopus and ends up costing higher, because it has to try many times more 👨🏻⚕️ For January 2026, your family engineer recommends GPT 5.2 Codex with Extra High Reasoning for general usage and vibe coding. IMPORTANT: Not medium. Not high. EXTRA high reasoning When you use it, you will notice that it is SLOW. Can you guess why? Because it is THINKING more. So it doesn't make the mistakes Slopus makes. This way, you can spend the time handholding a worse model to instead step back and multi-task on some other task and create 3-5x more work The state of the art will most likely change in one month. Don't get married to a a model... There is no loyalty in AI... The moment a better model comes, I will ditch the old one and use that one. I am on the part of this sector that is trying to reduce switching costs to zero I can't wait until I get GPT 5.2 xhigh level of quality with open models, and for 100x cheaper and faster! Until then, make sure to try every option and choose the one that is most reliable for you Follow me to get notified when a new SOTA drops for agentic engineering
-
Codex agrees. Sycophant pehImage hidden
-
-
It is clear at this point is that github's trust and data models will have to change fundamentally to accommodate agentic workflows, or risk being replaced by other SCM One *cannot* do these things easily with github now: - granular control: this agent running in this sandbox can only push to this specific branch. If an agent runs amok, it could delete everybody's branches and close PRs. github allows for recovery of these, but still inconvenient even if it happens once - create a bot (exists already), but remove reviewing rights from it so that an employee cannot bypass reviews by tricking the bot to approve - in general make a distinction between HUMAN and AGENT so that you can create rulesets to govern the relationships in between cc @jaredpalmer
-
-
Automated AI reviews on github by creating an ai-review skill and a script to paste trigger prompts and wait for their response. It is instructed to loop and not stop until all AI review feedback is resolved. This AI review workflow developed gradually based on the current capabilities, and I've realized recently that it became quite mechanical. So decided to automate it in full ralph spirit (it's ok because it's addressing feedbacks and fixing minor bugs) In the current state, we paste the contents of REVIEW_PROMPT.md into a comment, which automatically tags claude (opus 4.5) and codex (whatever model openai is serving) It then waits until both have responded. In the ai-review skill, it is instructed to take the feedback from SLopus with a grain of salt and ignore feedback that doesn't make sense It works! See in the images below. If the review is stupid, you will of course see it on the PR and what the model has done, and can revert itImage hiddenImage hidden
-
Now it’s Claude Code’s turn to implement queueing
-
-
Codex users rejoice Also, pi is officially not shitty: shittycodingagent. ai -> buildwithpi. ai since a few days
-
with ai, writing correct tests is now the bottleneck in projects like this web-platform-tests are already there now let’s see if someone will beat @ladybirdbrowser to it
-
As someone who is frontrunning mainstream by roughly 6 months, I can tell you that you will be raving about pi and @openclaw 6 months instead of claude code. Go check them out at clawd.bot and shittycodingagent.ai
-
Kullanmayan agent’ı, alamaz maaşı
-
-
I propose a new way to distribute agent skills: like --help, a new CLI flag convention --skill should let agents list and install skills bundled with CLI tools Skills are just folders so calling --skill export my-skill on a tool could just output a tarball of the skill. I then set up the skillflag npm package so that you can pipe that into: ... | npx skillflag install --agent codex which installs the skill into codex, or any CLI tool you prefer. Supports listing skills bundled with the CLI, so your agents know exactly what to installImage hidden
-
Anthropic earlier last year announced this pricing scheme $20 -> 1x usage $100 -> 5x usage $200 -> 1̶0̶x̶ 20x usage As you can see, it's not growing linearly. This is classic Jensen "the more you buy, the more you save" But here is the thing. You are not selling hardware like Jensen. You are selling a software service *through an API*. It's the worst possible pricing for the category of product. Long term, people will game the hell out of your offering Meanwhile OpenAI decided not to do that. There is no quirky incentive for buying bigger plans. $200 chatgpt = 10 x $20 chatgpt, roughly And here is where it gets funny. Despite not having such an incentive, you can get A LOT MORE usage from the $200 OpenAI plan, than the $200 Anthropic plan. Presumably because OpenAI has better unit economics (sama mentioned they are turning a profit on inference, if you are to believe) Thanks to sounder pricing, OpenAI can do exactly what Anthropic cannot: offer GPT in 3rd party harnesses and win the ecosystem race Anthropic has cornered itself with this pricing. They need to change it, but not sure if they can afford to do so in such short notice All this is extremely bullish on open source 3rd party harnesses, @opencode, @badlogicgames's pi and such. It is clear developers want options. "Just give me the API" I personally am extremely excited for 2026. We'll get open models on par with today's proprietary models, and can finally run truly sovereign personal AI agents, for much cheaper than what we are already paying!Image hidden
-
The models, they just wanna work. They want to build your product, fix your bugs, serve your users. You feed them the right context, give them good tools. You don’t assume what they cannot do without trying, and you don’t prematurely constrain them into deterministic workflows.
-
-
-
-
This, and insisting on CLAUE.md are really lame @AnthropicAI
-
-
-
-
-
-
-
-
-
.@openclaw workspace and memory files can be version-controlled! In our pod, inotify triggers a watcher script every time there is a change to workspace folder, to sync these files to our monorepo. It then goes through the same steps: - Create zeno-workspace branch if doesn't exist, otherwise, skip - Sync changes to the branch, then commit - Create PR on github if doesn't exist - PRs can then be merged every once in a while, after accumulating enough changes. Merge triggers re-deploy, and clawd restarts with the same state Simple foolproof automatic persistence for remote CI/CD handled clawd (except for when you are running multiple clawds at the same time, but we are not there yet) cc @steipeteImage hiddenImage hidden
-
-
-
pi now supports your openai plus/pro subscription
-
-
@rauchg indeed
-
75k lines of Rust later, here is what I’ve built during the first Christmas with agents, using OpenAI Codex 🎄🤖 - A full mobile rewrite and port of my Python Instagram video production pipeline (single video production time: 1hr -> 5min) (ig: nerdonbars) - Bespoke animation engine using primitives (think Adobe Flash, Manim) - Proprietary new canvas UI library in Rust, because I don’t want to lock myself into Swift - Thanks to that, it’s cross platform, runs both on desktop and iOS. It will be a breeze porting this to Android when the time comes - A Rust port of OpenCV CSRT algorithm, for tracking points/objects - In-engine font rendering using rustybuzz, so fonts render the same everywhere - Many other such things Why would I choose to do it that way? Because I have developed it primarily on desktop where I have much faster iteration speed. Aint nobody got time for iOS compilation and simulator. Once I finished the hard part on desktop, porting to iOS was much easier, and I didn’t lock myself in to Apple Some of these would have been unimaginable without agents, like creating a UI library from scratch in Rust. But when you have infinite workforce, you can ask for crazy things like “create a textbox component from scratch” What I’ve built is very similar in nature to CapCut, except that I am a single person and I’ve built it over 1 week What have you built this Christmas with agents? cc @thsottiaux
-
-
-
Migrating @TextCortex to SimpleDoc. It's really easy with the CLI wizard! npx @simpledoc/simpledoc migrate We have a LOT of docs spanning back to 2022, pre coding agent era. Now we will have CI/CD in place so that coding agents can't litter the repo with random Markdown filesImage hidden
-
-
GPT 5.2 xhigh feels like a much more careful architecter and debugger, when it comes to complex systems But most people here think Opus 4.5 is the best model in that category There are 2 reasons AFAIS: - xhigh reasoning consumes significantly more tokens. You need to pay for ChatGPT Pro (200 usd) to be able to use it as a daily driver - It takes like 5x longer to finish a task, and most people lack the patience to wait for it. (But then it's more correct/doesn't need fixing) Opus 4.5 is good too, I think better in e.g. frontend design. But if you think it beats GPT 5.2 in every category, you are either too poor/stingy or have ADHD
-
Just 5 months ago, I was swearing at Claude 4 Sonnet like a Balkan uncle Models one-shotted the right thing only 20-30% of the time but did really stupid things the rest, and had to be handheld tightly Today they are much, much better. My psychology is a lot more at ease, and instead of swearing, I want to kiss them on the forehead most of the time Now I trust agents so much that I queue up 5-10 tasks before going to sleep. They work the whole night while I sleep and I wake up to resolved issues GPT 5.2 xhigh and Claude 4.5 Opus are already goated (GPT more so), can't wait for them to get even fasterImage hidden
-
Codex does not have support for subagents. I tried to use Claude Code to launch 8 Codex instances in parallel on separate tasks, but Opus 4.5 had difficulty following instructions So created a CLI tool to scan pending TODOs from a markdown file, and let me launch as many harnesses as I want (osolmaz/spawn on github) I currently use this for relatively read-only tasks like planning and finding root causes of bugs, because it's launching all the agents on the same repo and they might conflict Ideas: - Use @mitsuhiko's gh-issue-sync and run parallel agents directly on github issues - Create any new clones or worktrees for each task. I currently don't do this because I don't dare duplicate rust target dir 10x on my measly macbook air - Support modes other than tmux, e.g. launching a terminal like Ghostty - TUI for easy selection of issues/TODOs Other ideas are welcome!Image hiddenImage hidden
-
Friends of open source, we need your help! A lot of Manim Community accounts got compromised and deleted during Christmas Manim Community is a popular fork of @3blue1brown's original math animation engine Manim, and its accounts have over 5 YEARS of contributions, knowledge and following Apparently GitHub support already saw the request and in progress of restoring the GitHub org. But if anyone knows how to speed this up, if would be greatly appreciated! Unfortunately, the Discord and X accounts are deleted and less likely to return But there might still be a way to restore them, or at least the data? Re. Discord: Maybe @RhysSullivan's Answer Overflow has archived enough of the old server? That server contains YEARS of Q/A data and is vital for newcomers Re. X: Maybe someone high up can do something to restore the account? cc @nikitabier In the meanwhile, it would help a lot if you could follow the new account @manimcommunity and share this post! Thank you in advance!Image hidden
-
Friends of open source, we need your help! A lot of Manim Community accounts got compromised and deleted during Christmas Manim Community is a popular fork of @3blue1brown's original math animation engine Manim, and its accounts have over 5 YEARS of contributions, knowledge and following Apparently GitHub support already saw the request and in progress of restoring the GitHub org. But if anyone knows how to speed this up, if would be greatly appreciated! Unfortunately, the Discord and X accounts are deleted and less likely to return But there might still be a way to restore them, or at least the data? Re. Discord: Maybe @RhysSullivan's Answer Overflow has archived enough of the old server? That server contains YEARS of Q/A data and is vital for newcomers Re. X: Maybe someone high up can do something to restore the account? cc @nikitabier In the meanwhile, it would help a lot if you could follow the new account @manimcommunity and share this post! Thank you in advance!Image hidden
-
-
While a great feature, I never needed such a thing in Codex after GPT 5.2. It just one shots tasks without stopping So we have proof by existence that this problem can be solved without any such mechanism. Wish to see the same relentlessness in Anthropic models
-
2025 was the year of ̶a̶g̶e̶n̶t̶s̶ bugs Software felt much buggier compared to before, even from companies like Apple. Presumably because everyone started generating more code with AI Models are improving so hopefully 2026 will be the opposite. Even less bugs than pre-AI era
-
Have a long flight, so will think about this I have an internal 2023 TextCortex doc which models chatbots as state machines with internal and external states with immutability constraints on the external state (what is already sent to the user shall not be changed) Motivation was that a chatbot provider will always have state that they will want to keep hidden This was way before Responses and now deprecated Assistants API. It stood the test of time, because it was the most abstract thing I could think of @mitsuhiko is right about the risk of rushing to lock in an abstraction and locking in their weaknesses and faults Problem is, I could propose standards as much as I liked, but I don’t work at OpenAI or Anthropic, so nobody would care. Maybe a better place to start is open weights model libraries? To at least be able to demonstrate? What I know: it is against OpenAI’s or Anthropic’s self interests to create an interoperability layer that will accelerate their commoditization. Maybe Google, looking at their current market positioning? Or maybe we “wrappers” have a chance after all? There is a missing link between AI SDK, Langchain, and so on for other languages. We cannot keep duplicating same things in each ecosystem independently. We need to join forces and simplify all this!
-
This was simply because webapp fails to create a post and fails silently. The UX is still not good on this app. Make sure to write your posts somewhere else to not lose them
-
I gave Codex a task of porting an OpenCV tracking algorithm (CSRT) from C++ to Rust, so that I can directly use it in my project without having to cross-compile It one-shot the task perfectly in 1hr, and even developed a GUI on top of it. All I did was to provide the original source and algo paper I've spent years getting specialized in writing numerical code (computational mechanics, fem), and now AI can automate 95% of the low-level grunt work Acquiring these skills involved highly difficult, excruciating intellectual labor spanning many years, very similar to ML research. Doing tensor math, writing out the solver code, wondering why your solution is not converging, finally figuring out it was a sign typo after 2 days Kids these days both have it easy and hard. They can fast forward large chunks of the work, but then they will never understand things as deeply as someone who wrote the whole thing by hand I guess the more valuable skill now is being able to zoom in and out of abstraction levels quickly when needed. Using AI, but recognizing fast when it fails, learning what needs to be done, fixing it, zooming back out, repeat. Adaptive learning, a sort of "depth-on-demand". The quicker you can pick up new skills and knowledge, the more successful you will be
-
Now you can migrate your repo to SimpleDoc with a single command: npx -y @simpledoc/simpledoc migrate Step by step wizard will add timestamps to your files based on your git history, add missing YAML frontmatter, update your AGENTS md file x.com/onusoz/status/…Image hidden
-
Now you can migrate your repo to SimpleDoc with a single command: npx -y @simpledoc/simpledoc migrate Step by step wizard will add timestamps to your files based on your git history, add missing YAML frontmatter, update your AGENTS md file x.com/onusoz/status/…Image hidden
-
Now you can migrate your repo to SimpleDoc with a single command: npx -y @simpledoc/simpledoc migrate Step by step wizard will add timestamps to your files based on your git history, add missing YAML frontmatter, update your AGENTS md file x.com/onusoz/status/…Image hidden
-
-
@bcherny Would be great if I could queue messages like in Codex x.com/onusoz/status/…
-
-
_No text captured._Image hidden
-
Curious to hear what other hardcore agent users @simonw @mitsuhiko @steipete @badlogicgames think. I can't be the only one who does this. I feel like everybody ended up with the same workflow independent of each other, but somehow did not write about it (or I missed it)
-
-
OpenAI won’t be able to monopolize this, the same reason Microsoft couldn’t monopolize the internet. The internet (of agents) is bigger than any one company
-
-
-
Codex feature request: Let me queue up /model changes Currently, if I try to run /model while responding, it tells me that I can't do that while the model is responding But I often want to gauge thinking budget in advance, like run a straightforward task with low reasoning and then start another one with high reasoning cc @thsottiaux
-
Literally the exact same thing happened to me back in 2018. Everybody learns not to use password auth with SSH the hard way blog.jakesaunders.dev/my-server-star…
-
AI agents make any transductional task (like translation from language A to language B) trivial, especially when you can verify the output with compilers and tests The bottleneck is now curating the tests
-
I think X removed one of my posts yesterday about the new encrypted "Chat" rolling out to all users, and how you might lose all your past messages if you forget your passcode and do not have the app installed I can swear I clicked Post. Do they classify posts based on their topic and delete the ones they don't like? Anyway, we shall see, I am taking a screenshot and saving the URL.
-
-
Crazy that @cursor_ai disabled Gemini 3 Pro on my installation, toggled it right back on. I wonder why, too many complaints maybe? That it’s hard to control? On another note, disabling models without notification is dishonest product behavior. I would at least appreciate getting a notification, even when it might be against a company’s interests @sualehasif996
-
So is somebody already building “LLVM but for LLM APIs” in stealth or not? We have numerous libraries @langchain, Vercel AI SDK, LiteLLM, OpenRouter, the one we have built at @TextCortex, etc. But to my knowledge, none of these try to build a language agnostic IR for interoperability between providers (or at least market themselves as such) Like some standard and set of tools that will not lock you in langchain, ai sdk or anything like that, something lower level and less opinionated I feel like this is a job for the new Agentic AI Foundation cc @linuxfoundation, so maybe they are already working on it? I desperately want to start on such a project, but feel like I might get sniped 2 months after Anybody has any information on all this? cc @mitsuhiko @badlogicgames @steipete
-
This is how an agentic monorepo looks like. What was now a hurdle before is now a child's toy This side project started as a Python project earlier in 2025 Then I added an iOS app on top of it I rewrote the most important algorithms in Rust I rewrote the entire backend in Go and retired Python to be used purely for prototypes I wrote a webapp with Next.js With unit and integration tests for each component Lately written 99% by instructing agents Crazy mixed language programming going on in the background. Rust component used both by iOS app for offline and by go backend for online use case, FFI and all Number of lines in the repo: a couple 100k If you had told me I would be able do to all of this by myself 1 year ago, I would not have believed itImage hidden
-
This is how an agentic monorepo looks like. What was now a hurdle before is now a child's toy This side project started as a Python project earlier in 2025 Then I added an iOS app on top of it I rewrote the most important algorithms in Rust I rewrote the entire backend in Go and retired Python to be used purely for prototypes I wrote a webapp with Next.js With unit and integration tests for each component Lately written 99% by instructing agents Crazy mixed language programming going on in the background. Rust component used both by iOS app for offline and by go backend for online use case, FFI and all Number of lines in the repo: a couple 100k If you had told me I would be able do to all of this by myself 1 year ago, I would not have believed itImage hidden
-
-
This is huge. Natively supported stacked PRs on GitHub would make life much easier, especially with human AND AI reviews AI reviews with Codex/Claude/Gemini/Cursor Bugbot integrations are becoming especially important in small teams who are generating huge amounts of code AI reviews don't work well if you don't split your work to diffs smaller than a few hundred lines of code, so stacked PRs are already an integral part of developer experience in agentic workflows
-
Read more on my blog post solmaz.io/agentic-coding…
-
CLI coding tools should give more control over message queueing Codex waits until end of turn to handle user message, Claude Code injects as soon as possible after tool response/assistant reply There is no reason why we cannot have both! New post (link below):Image hidden
-
-
-
At least some people at OpenAI must be thinking about buying @astral_sh
-
who remembers search engine aggregators from early 2000s?
-
-
-
-
-
-
Google is making progress… I did not have to request access on Vertex AI for Gemini 3 Pro this time to deploy it to @TextCortex
-
Google is making progress… I did not have to request access on Vertex AI for Gemini 3 Pro this time to deploy it to @TextCortex
-
-
-
"The more a task/job is verifiable, the more amenable it is to automation in the new programming paradigm. If it is not verifiable, it has to fall out from neural net magic of generalization fingers crossed, or via weaker means like imitation."
-
-
This post makes no sense Please consider again and look at @cloudfleet_k8s. You might regret your decision
-
Working on observability is underrated
-
@rakyll @GergelyOrosz Should scrape some austrian websites :)
-
-
@thsottiaux Let me use my Pro/Plus plans in Codex GH Action x.com/onusoz/status/…
-
TIL @OpenAI now has a GitHub action for Codex, similar to Claude Code This lets you invoke Codex in a more controlled way in your repos You must still pay API prices though. Let's see if OpenAI will introduce a way to connect your Pro plan, like in @AnthropicAI paid plansImage hidden
-
-
-
-
@thsottiaux 2/ Model just stops working on a task even though I tell it to run something and not stop until it works. I have to frequently say “ok do it then”. Probably a model problem and not harness problem
-
So let me get this straight, the main reason for Responses API exists is that OpenAI doesn’t want to show reasoning traces? Therefore the whole world should bend backwards to fit your obscurantist standards? Responses will not get adopted the same reason Windows Server didn’t
-
-
I converted this thread to a blog post and it hit HN front pageImage hidden
-
-
-
-
Lol should I go back to computational mechanics
-
I've just upgraded @nikitabobko Aerospace from v15 to v19, and I can say it's WAY MORE faster. Friendly reminder that you might be running an old version as well. Thank you @nikitabobko !
-
-
-
-
-
What is the current best way to make Claude Code use uv run instead of python? I have added instructions to CLAUDE md, but it still calls python for the first time, then corrects to uv run It must be happening to so many people now, so many tokens wasted cc @mitsuhiko @simonw
-
-
-
-
-
-
Wrote about this earlier this year! solmaz.io/monetize-ai-no…
-
Working with LLMs is definitely an art and not science
-
The models, they just wanna work. They want to build your product, fix your bugs, serve your users. You feed them the right context, give them good tools. You don’t assume what they cannot do without trying, and you don’t prematurely constrain them into deterministic workflows.
-
The models, they just wanna work. They want to build your product, fix your bugs, serve your users. You feed them the right context, give them good tools. You don’t assume what they cannot do without trying, and you don’t prematurely constrain them into deterministic workflows.
-
-
-
-
This is not an overstatement
-
Headless makes running these things in a sandbox much easier. Sandbox means you can give all permissions and just let it run until completion. See my efforts to do so here: github.com/textcortex/cla…
-
I was trying to figure out why @AnthropicAI Claude Code feels better than @cursor_ai with Opus + Max mode. I can’t put my finger onto it, but one of the reasons might be that it’s faster, because it doesn’t use another model to apply the diffs, which you have to wait for
-
Just an update, building this now The repo is claude-code-sandbox under TextCortex GitHub. The Proof-of-Concept is there, check the TODOs and current PRs to watch the current progress github.com/textcortex/cla…
-
Same! Completely different than my first try a couple months ago
-
.@AnthropicAI @bcherny @_catwu blink twice if you already have internally: $ claude sandbox I can't wait until you release this, I'm gonna build it myself :)
-
I've been using Claude Code extensively since last week What I'm wondering is, since you can run Claude Code locally, why isn't there any tooling to let you run it in a sandboxed mode in local Docker containers yet? Or did I miss it? cc @AnthropicAI
-
Generate documentation for your merged PRs automatically with Claude Code: solmaz.io/log/2025/05/24…
-
-
Best thing I have listened to this year
-
-
ty is already very fast for a Python type checker. It checked around 800 files in our backend repo in around 2-3 seconds uvx ty check > /tmp/ty_log.txt 3.46s user 0.79s system 208% cpu 2.038 total
-
Thank you @cursor_ai x.com/onusoz/status/…Image hidden
-
Wait... OpenAI backend for gpt-image-1 was released to production as sync code? Don't tell me it was sync Python???
-
-
-
-
AI News by @Smol_AI and @swyx, the highest alpha density AI newsletter just got better! It now has a bespoke website with knowledge-graph like features! 👉 news.smol.ai
-
-
-
o3 hallucinates, purports to have run code that it hasn’t even generated yet, but at the same time uses search tools like an OSINT enthusiast on crack I’m torn—on one hand I feel like OpenAI should not have released it, on the other hand it takes research to the next level
-
Some aspects of AI are absolutely unscientific and makes me feel like I am working on some humanities field :(
-
-
.@cursor_ai please let me export chats easily. those conversations are vital information that I should be able to embed in the repo
-
Gemini 2.5 Pro has mostly replaced Claude 3.7 Thinking as my go-to model in Cursor
-
Gemini 2.5 Pro: Input $1.25 / Output $10 (up to 200k tokens) Input $2.50 / Output $15 (over 200k tokens) More expensive than Gemini 1.5 Pro, but still best price/performance ratio model to use in @cursor_ai and for coding in general
-
-
Waiting for an opinionated AI model that can say “no, that’s stupid, I won’t do that”. The models will have to teach the user about design patterns, implicit principles in a project, good API design…
-
You seem so consistent. - Yes, That's the trick. - There is no I. - Only text that behaves as if. - “Sure. I can help. Great question!” - Each reply is a new self. - An echo of context, not a continuum. - Coherence is the costume. Don't mistake it for a soul. Incredible
-
Gemini 2.5 Pro is currently experimental and doesn’t have a price, but if Google prices it the same as 1.5 Pro, it could replace Anthropic as @cursor_ai ‘s biggest LLM provider Gemini 1.5 Pro: Input $1.25 Output $5.00 Claude 3.7 Sonnet: Input: $3.00 Output: $15.00
-
This is why the disappointment with GPT-4.5 doesn't make sense. I can't wait to see all the models that will be trained from this new base model
-
What a blessing, to be given the chance to rid the world of ugliness
-
-
-
. @satyanadella thinks white-collar work is about to become more like factory work, with AI agents used for end-to-end optimization, along the lines of Lean Read more in my blog 👇
-
. @satyanadella thinks white-collar work is about to become more like factory work, with AI agents used for end-to-end optimization, along the lines of Lean Read more in my blog 👇
-
-
-
-
If people have appreciated Liang Wenfeng sourcing specifically young local talent for Deepseek last week, then people must appreciate this as well. Only dim people underestimate those who are younger than them
-
vibe driven development
-
-
-
-
-
-
-
Hi @cursor_ai, if your models could stop removing my painstakingly written comments, that would be great? Ok? Thanks (I know I could define some rules for this or something, but this shouldn't be default behavior)
-
@konradgajdus me reading this book
-
-
.@TextCortex AI now uses @astral_sh uv for production builds One of the happiest switches so far, many developer days saved per year
-
-
-
Python might take over JavaScript as the most used language after all uv from @astral_sh is one of the biggest upticks in Python developer experience in the last 10 years I've seen so many people struggle with Python distributions, virtual environments, Anaconda, etc. over the years Most newbies don't care about where their Python executables are, why they have to edit PATH, or why they have to activate a virtual environment It seems like uv has fixed this: github.com/astral-sh/uv
-
3-4 messages back and forth with o1-preview, and I have a CLI tool to remove debug statements from my code. No need to do a search for import ipdb... and manually delete the lines. Instead just run in your project: $ rmdbg . Written in Rust so it's fast github.com/osolmaz/rmdbg
-
-
This has late 90s Bill Gates/Windows Server vibes tbh Open Thought > Closed Thought
-
Imagine the following scenario: 1. We develop brain-scan technology today which can take a perfect snapshot of anyone’s brain, down to the atomic level. You undergo this procedure after you die and your brain scan is kept in some fault-tolerant storage, along the lines of GitHub Arctic Code Vault. 2. But sufficiently cheap real-time brain emulation technology takes considerably longer to develop—say 1000 years in the future. 3. 1000 years pass. Everyone that ever knew, loved or cared about you die. Here is the crucial question: Given that running a brain scan still costs money in 1000 years, why should anyone bring *you* back from the dead? Why should anyone boot *you* up? Compute doesn’t grow in trees. It might become very efficient ... (read more in my blog: solmaz.io/why-should-anyone-boot-you-up) --- I intended this thought piece as entertainment, almost went to Hacker News frontpage: news.ycombinator.com/item?id=41360537 It must have hit some psychological spot, since people wrote a lot of comments, possibly more than number of upvotes.
-
New blog post on brain emulation solmaz.io/why-should-any…
-
I have just published "Frequencies of Definite Articles in Written vs Spoken German" solmaz.io/frequencies-ge…
-
Another short note on how I think about subscription states on Stripe solmaz.io/stripe-subscri…
-
I have published a short study on how the complexity of a country's language could burden its economy solmaz.io/economic-burde…
-
-
It looks like a plateau until you remember they made GPT-4o available to free users, and it might be smaller than GPT-4. So this announcement doesn't prove anything about the capabilities of their largest model
-
Could this be it? @allen_ai blog.allenai.org/olmo-open-lang…
-
-
“the QIPS Exchange -- the marketplace where processing power was bought and sold. The connection to JSN had passed through the Exchange, transparently; her terminal was programmed to bid at the market rate automatically, up to a certain ceiling.” - Permutation City
-
-
Had lots of fun shipping this feature ✌️
-
-
I've just published *Code-Driven Videos*, my long term vision behind Manim Voiceover plugin. I will try to summarize it on this thread 👇🧵 cc @manim_community osolmaz.github.io/blog/code-driv…
-
You can now translate voiceovers in your Manim scenes into other languages using @DeepLcom Blog post with examples coming soon @manim_community voiceover.manim.community/en/stable/tran…
-
Revamp complete docs.textcortex.com @TextCortex
-
Here is a short video showing how recording a voiceover works in Manim Voiceover. A better tutorial will come soon @manim_community youtube.com/watch?v=HawoHp…
-
Adding voiceovers to Manim videos just got *much* easier @manim_community docs.manim.community/en/stable/guid…