Entries for August 2026

me: I want to publish in arXiv mom: we have arXiv at home arXiv at home: jokes aside, my astro blog can now render markdown/astro posts as if they were latex papers because latex unfortunately has unreasonable effectiveness in convincing people that an idea is important, even though it may not be my random shower thought got 200k views earlier, whereas what I really thought was a big deal dwindled so maybe this will help it a second time useful for literally everyone who works on optimizing inference Theoretical Upper Bounds for LLM Throughput (wip, shoot corrections in the replies):

I remember being excited about Fable that we finally have a chungus model like GPT 4.5 that can write great prose Oh how naive I was. RLVR or something else in posttraining introduced since ruins models' ability to speak in an understandable tone I have just STRUGGLED trying to make Fable write an abstract, and had to write it myself at the end. We have AGI, but it cannot write an abstract :( I wonder whether the labs are already planning to fix this? Like checking for unreadable writing can easily be made deterministic, with a readability score and such. Though adding a reward over that would probably make things worse, so I'm not sure Like some of these issues must be easy to fix. Consider "sentence parade", where each sentence in a paragraph is completely detached from each other. Like "A does B. C is D. E does F" and so on. For example: > The run reaches 30% of the ceiling. Decode reached 77% to 86% of its bound. The gap says software is the limit. Sustained FLOP/s sits below the bracket. Here is a vibeslopped script which detects such cases with spacy, with surprisingly high precision: github.com/osolmaz/ai-smell/blob/main/analyze_c… I was reading @ben_burtenshaw's preview of his Post-training book---might be a fun weekend project, training a small model for increasing readability

glm 5.3 flash at an antirez-style asymmetric q2 quant could plausibly fit on a single dgx spark at 100~110 GB but with 18b active params vs ds4 flash’s 13b, theoretical decode throughput is only ~70% of that of ds4 flash on the same hardware that is 28 tok/s ds4 vs 19~20 tok/s glm, without speculative decoding, at 32k token filled context afaik it hasn't been quantized in that style yet, but if it were, we would expect such throughput ratio between those models

Qwen 3.8 Flash-Next will have to be quantized down to 4 bit precision or lower in order to fit a DGX Spark and leave some leeway Excited to see the performance on the Spark

this now supports auto syncing skills to .claude/ folder as well

wake up babe, it is 2026 and we can only afford to lease computers now also, holy mother of 1.2 TB/s bandwidth 🥵

*autoplan* This is one of my most used workflows in pi now While working with AI agents, there is a mechanical process by which I mine the agent for ideas. This reduces an open-ended feature design or bugfixing problem to a multiple choice question Basically, I keep asking paraphrases of the question "is this the best design?" 2-3 times, and then make the agent list them out, with a preference for practicality and simplicity "Is this the most elegant and long-term production ready solution?" "Is this the holy grail?" And then a decision gate which makes the model list all options while recommending a certain one, with a preference for practicality For example, while developing a plugin for pi or openclaw, asking the holy grail often causes the model to suggest changing the plugin/extension API like "The holy grail would be for pi to implement such an such API". The decision gate helps curb such stupid ideas The good thing about this workflow is, I can just automate typing all those mining prompts, and only do the deciding after the workflow finishes Caveat: This is not foolproof. I still reject all options, propose other ones, or run the workflow multiple times until I get what I want. But this helps reduce a ton of prompting to just "autoplan this" for me I am curious: When you try this, does it give you high quality answers/designs? And if not, what should change to improve it? To try it out: Install osolmaz/pi-workflows and then when you need to design something or fix a bug, just say "autoplan this". The skill should be picked up automatically Let it finish. It will give you a summary. When you choose an option, ask it to elaborate it with more details Repo:

People want to migrate from CLAUDE md so much, a blog post I wrote on it last year is still one of my most visited posts I also wrote a service that auto symlinks when Claude uses a directory that has AGENTS md. Basically seamless, automatically gitignored symlinks. I use Claude with AGENTS md and had forgotten that this problem existed Repo: github.com/osolmaz/claude-md-symlinker

Pi folk, there is an issue with compaction in pi that might drain your codex limits faster There is a 272k token limit after which the API charges 2x for input and 1.5x for output tokens Issue: github.com/earendil-works/pi/issues/6879 @cancelik's extension is a workaround until this is solved upstream in pi:

@frankekn @Zai_org hmmmm x.com/di_zhang_fdu/s…

puny 40k volume but polymarket seems to think ox alpha is GLM by @Zai_org maybe we were thinking too complicated after all

I was not expecting this to receive so much attention! There was a lot of constructive feedback to what was basically an outcry after battling the complexity of the benchmarking space for 2 months straight I agree with the points that models are post-trained to be efficient in a certain harness, so an objective measurement must use that But I also see that it is very convenient for proprietary model vendors for everyone to accept that the model is inseparable from the harness The main issue is that model population (as well as harness) is about to explode, and there must be an effort to standardize things. That should be done by an industry-wide consortium and/or a neutral third party I am not going to propose a harness in this post, but just share an idea that, if we had agreed on one, how we would reconcile models' baselines across different harnesses Idea: We could run each benchmark on at least 2 harnesses: Let S be the score of the model on a given benchmark using the *standardized test harness* and N be the score of the model over the *native harness* which the vendor post-trained the model with Define "Standard Harness Sensitivity"[*] as: SHS = (N - S) / (N + S) Which gives normalized score between -1 and 1 that measures how sensitive a model is to being put in a different harness other than the recommended one. 0 means that the model is neutral to which harness it's in (at least for the given 2) 1 means it is much worse outside the native harness, for that benchmark -1 means for some reason, it functions much better in the standard harness than the native harness, which might imply that the vendor messed something up So just by running a benchmark 2 times and reporting 3 quantities S/N/SHS, you can get a lot of info: - How well a model would perform if you were to use the native harness, - how well it would perform if you were to use a more simple and neutral harness, - and how sensitive it is to change of harness Running benchmarks is EXPENSIVE, and the cheaper we can extract this signal (i.e. with 2 harnesses instead of N), the better --- [*] I would call it just Harness Sensitivity, but apparently that term is already used

sol really loves to use words like envelope, capsule, receipt

I have performed a distillation attack on my PERSONAL TRAINER Since more than 2.5 years, I have recorded over 300 of our sessions, each up to 90 minutes long Why would I do that? Because he is a kinesthetic genius from the Caucasus who developed his own training doctrine He loves helping people become fitter and freer in their bodies, counting every rep and cheering sincerely when they unlock a new skill And now both he and I would like to share it with the rest of the world To begin with, I used Whisper+pyannotate to automatically transcribe all our sessions and used LLMs to extract all the knowledge He taught me over 100 movements and spoke about over 100 topics over the course of 2.5 years You can browse and read all of them here, for free: mirzamovement.com/wiki/ But again, why all this? 1.5 years ago, I wrote an article called "Our muscles will atrophy as we climb the Kardashev Scale" which went to the Hacker News front page It was sort of a meme, but it reflected a reality that I felt very deeply: Having a sedentary job which only requires me to use my brain is shortening my life and reducing my life quality by a lot Not only mine, but of hundreds of millions, soon billions of people who will spend hours just talking to AIs whole day for their careers It is also a fact that personal physical education and training is close to nonexistent for the average person There are billions of people who could live much healthier lives, just by exercising in the room they already have. But they just don't know how. And there is no system to teach them how, effectively at scale 1 on 1 personal training should be universal and free, with AI We must have reps and sets too cheap to meter An expert AI PT that sees and corrects your every move, that counts for you. Created by people that care, like us. Unlike a franchise gym that just wants to get your subscription and sell you protein powder Once everybody can taste good AI personal training for free, the demand for human personal trainers will explode and whole fitness industry will see unprecedented growth That is one of my life's goals, and I will make that happen one day. Today is the first step My trainer is Orkhan Mirza, a.k.a. @JashinJashua here and on Instagram, and I am nerdonbars on Instagram The wiki and more are available here: mirzamovement.com Let me know what you think!

We need to normalize measuring and judging models against a standardized test harness "Oh but model X performs best in their own proprietary harness" I could not care less. When I take exams, I go to the standardized classroom, get the standardized pencil and exam sheet, and have to solve it under 2 hours This system arose because we have a LOT of people to test Guess what? We now have a LOT of models, and they are multiplying by the day "Oh but model X performs substantially better in ARC-AGI-3 with a custom harness" I don't care... Then imbue model X with enough knowledge so that it can reconstruct that harness on the spot The main harness could be mini-swe-agent, terminus 2, vanilla pi or something along those lines It needs to be simple, and stay roughly the same over time There is already too much complexity in the benchmarking space right now, and I feel like not enough people are putting their feet down to cut away some part of it

This actually makes the most sense to me from what I've read till now. Wouldn't be surprised if it were Elon's idea to use the name "Ox Alpha" as bait

my @pidotdev extensions give me so much joy. so much tedium that I automated away, that I know are guaranteed to be performed the way I want, unlike with pure skill files

Who is actually training a model to come up with better headlines/thumbnails? This can be end to end automated to some degree

What a time to be alive @codex_reset

This is your daily reminder that you can just vibe your dream terminal dashboards with @ratatui_rs and it just works amazingly

Once nice thing with workflow graphs is to be able to compose them I stitched `autoplan` + `autoimplement` workflows into `monitor` in an "issue detected" path: monitor -> wait 30min -> issue detected -> autoplan -> autoimplement -> monitor Previously, monitor would just stop with "blocked" on trivial issues. Now, the workflow deterministically challenges the model whether it actually is a blocker (it literally asks, "are you really sure if it's a blocker") If the model is like "nah bro, it's actually something I could fix", then it puts the model into this path The good thing is, I can compose and nest workflows arbitrarily without code duplication I generally call `autoplan` and `autoimplement` separately. I prefer to read the plans before I hit implement But by the time I hit `monitor` on a long running job, I expect something to finish by itself (I have given all the information that the system needs, and any new issues should be trivial) In the screenshot below, the model went into this path and unblocked itself automatically Repo (still highly experimental, breaking changes will happen):

We live in a simulation

God save my soul I might have built quite a bit of a rube goldberg machine here in pi My "autoimplement" skill of 6 months is now an actual graph based workflow I even have "autoplan". It automates the inquisition I make the model perform on itself. It creates the plan, and pings me on telegram whether or not to move forward with it The idea is to have reactive agents which auto-react to events, plan, ask for approval and then fix things as autonomously as possible The good thing is, I can now enforce the model programmatically not to chase P2 and P3 review issues ad infinitum Same for long running CI. I can now automatically detect inefficient CI that takes longer than 5~10 minutes and optimize it automatically as well Let me know what works for you and what breaks (wip):

Before someone says it, I know github.com/earendil-works… exists. But it doesn't give me a CLI I can just run

A pi-based alternative to codex review, for your auto-review loops It is like codex review, but it works with all the models that pi supports, including gpt 5.6 github.com/osolmaz/pi-rev…

It seems my tweet ragebaited some unintentionally (best kind of ragebait, with plausible deniability) My point here was that, you would be sending a compressed version of human knowledge back in time Having compressed such high intelligence per parameter, Qwen 3.8 27b can be mined for a lot more knowledge than a 20gb fraction of Wikipedia Here is how that would have worked

POV: You've automated monitoring your agents I did a lot of quality-of-life improvements to osolmaz/pi-workflows. The pi widget now only shows a compressed list instead of the full graph If you use herdr, you can still open the graph viewer in a side pane with Ctrl+Shift+R. Your pi extensions can be herdr aware too! This means pi-workflows is also a @herdrdev plugin now, alongside @pidotdev Monitor is currently my most used workflow. Let's the model submit any long-running task progress and calculate ETA automatically Just say "monitor every 30 minutes and resolve any issues autonomously", or even just "monitor". The shipped monitor skill teaches the model how to run this Repo:

before anyone thinks anything, I am using codex's remote compaction in pi btw

Qwen 3.8 27b weights are probably the most civilization-changing 20gb of data published, until now Imagine sending those 20gbs back in time, to 2006

The only reason I'm using Sol at this point is that it's subsidized and fast tbh

How insane is tailscale/wireguard Discord is blocked in turkey, so I just set my spark as an exit node. And I can use it as VPN without any further config

People are going nuts over qwen 3.8 27b and it's justified I'm personally waiting for qwen3.8-35b-a3b. When released, it will be the best model to run openclaw on the dgx spark

.@odd_joel I want to install Moshi hooks but they are closed. They have security implications, so I would rather install sth open source on my machine Any chance you might open source the hooks?

Qwen 3.8 27b just scored 4% higher than GLM 5.2 in a non-trivial private benchmark I am working on Not sure if there is an issue with the benchmark or the new 27b is that good

😎 local.ai/onur/invite

A similar thing can be done with codex session compaction summaries as well btw I had verified the other day that a compaction summary created by one account can be used by another account if it has the encrypted blob ie. they were not keyed/guarded with your account id

This type of estimations are useful not only for local but for all inference providers Get a ballpark of max possible throughput for a model, directly calculate your revenue

mfw trying not to think about the 100 models that dropped this week I’ll be offline 1 week to touch grass and reset I’ve literally not given any break since claude code came out lol let’s see if I can resist the urge to check twitter

I've felt the lack of a certain feature of codex desktop app since I went back to the CLI: scheduled tasks Codex desktop app can keep track of a task until it is properly finished. It's basically cron. And for some reason, codex CLI still doesn't have it. Codex app acts as a shared runtime, and for some reason, certain features don't work without it, even though they could... there is no reason for openai to not use a background process So I got bored of waiting, and decided to build my own in @pidotdev But I realized, I could do much more than a simple cron job, with my recently upgraded osolmaz/pi-workflows extension A cron job is a loop after all. Being a loop, I can represent it as a workflow graph So I created a built in `monitor` workflow to mimic cron behavior. The agent is forced into a loop where it re-checks a very long-running job every 1 hour, and it is instructed to autonomously correct it and fix any bugs if any are encountered The same functionality can be achieved by iamwrm/pi-unified-exec as well, which implements codex-like auto-forking exec behavior. But there is a chance the model messes up exec, or does not re-arm the next sleep() properly once one of them exits My monitor workflow on the other hand is deterministic. I can make the agent loop infinitely, and there is nothing the agent can do to evade the task. I just ask the agent to monitor something, and it starts it automatically This lets me just fire off week-long jobs, and forget about it! It even survives codex usage depletion, by auto-recovering once my quota resets Oh also, @ratatui_rs is a delight! I created piw, a viewer for my ongoing pi workflows. I just type piw, and can see the current state, or play back the finished ones

Apparently this went to HN front page briefly last night, so sharing it here as well YOLO safely with your agents 🤖 Give your GitHub/Hugging Face accounts + ability to run sudo safely to your agent. No need to create an agent account, or clickops policies on GitHub etc. Give it merge access to repo X for 5 minutes, 30 minutes, 1 time, 100 times, anything... Then give it unlimited access to repo Y forever. Other repos stay untouched. Complete flexibility that GitHub policies actually cannot give you due to the way that they are designed You can get notifications through telegram, and can approve its requests No need to pay $4 to GitHub if you simply want protection against force push. unYOLO blocks force pushes by default, unless you explicitly allow to Your exfiltratable, internet-accessing agent/claw gets its own Linux/Mac account, and has to access these services through unYOLO The video shows how it works! Visit:

This. LPDDR chips are cheaper to produce and run GDDR/HBM will likely keep being more expensive Most consumer GPUs will converge on a GB10 like form factor As much as us hobbyists love to project this ideal of running a GPU cluster at home, most working people will prefer smaller form factors, and will not want to pay hundreds of $$$ in electricity bills every month DGX Spark/GB10 runs at around 90-150 Watts RTX Pro 6000 runs at 600 Watts FOR THE GPU ALONE, and can cost 3-5x more than GB10. Despite having 25% less memory capacity than GB10... Looking at this, LPDDR will be orders of magnitude more commonplace at home Architectures will develop accordingly. Future local AI will be dominated by MoE and similar architectures which leverage mid-sized models with smaller number of active parameters That is why Qwen3.x-35B-A3B is a more useful model on the Spark than Qwen3.x-27B, despite the latter being a better model. Same for Gemma I can run A3B at 60 decode tok/s single session or 6x20 decode tok/s in parallel, whereas 27b only reaches 1/3rd of that Future of local AI is DDR/LPDDR and MoE/adjacent architectures, for the average person

urgh typo, I meant 35-40 tok/s

Apparently Alibaba did all this work, published a paper, but did not create a public backup of repos used in the review tasks arxiv.org/html/2601.19494v2 Then keycloak and nodejs repos got force pushed, so the commits for 6 of the tasks got lost :( I recovered 4 of them, but 2 commits are still missing: keycloak/keycloak#35645 460f8008f86d3fa8f62da63e26d8bdc306af60b2 nodejs/node#56185 b2255442712cb6db83d112deb6ba61197d06a5f3 Would anyone happen to have them backed up locally or somewhere in a fork?

I have made an update to my theoretical upper bound calculation to also predict prefill speed Prefill relaxes the assumption we make for decode, that it is only be memory bottlenecked. So prefill can be both compute or memory bottlenecked. I use the FLOP limits reported by hardware producers for the estimates: These estimates will also be available in ourmodels.cc for indexed model and hardware in a couple days, once a long running job finishes Blog post: solmaz.io/llm-performance-upper-bounds

You can use the calculator in localmaxxing.com/en now to predict theoretical upper bounds from solmaz.io/llm-performance-upper-bounds 🙌 A little note, pay attention to the speculative decoding coefficient rho. If an upper bound is calculated using rho=1, then dspark might push it well above that limit. For example, the formulation predicts an upper bound of 28 decode tok/s, but dspark pushes it up to 35-50 tok/s in github.com/0xSero/deepseek-v4-flash-0731-spark-… by @0xSero I am curious whether we will observe a phenomenological law between real-life engine performance and the upper bound, like real-life performance maxes out at e.g 80% of the upper bound in most cases That would be very useful! Then we would be able to tell when something is wrong with an inference engine, if it doesn't reach at least 80% of the upper bound, ignoring spec. decoding

Rules that current models cannot follow with skill, i.e. a single prompt: Chasing P2 and P3 errors: I have a rule in my autoimplement skill to stop reviewing once the last round of review only generates P2 errors or less. A considerable % of the time, the model just goes on an adventure addressing all the issues it can find Running checks and CI efficiently: I have rules like "commit and merge opportunistically, make sure to not wait for irrelevant tests". But it keeps waiting for 30 minutes of CI before merging, in every commit for every little fix, every time Those are the cases where graph workflows are needed. One can deterministically enforce the model not to take longer than N minutes reviewing, or fixing CI Or post to the agent to hurry up when it is taking too long (which I believe might be effective, since RL envs also have time constraints)

I'm running some private benchmarks on @liquidai's LFM 2.5 2.6B, and if my results are correct, we might have a new champion for <10b category Scores significantly higher than gemma 4 e4b, which is 3-4x its size

Anyone else use Alibaba’s AACR bench for code review evaluation? github.com/alibaba/aacr-bench I am adapting it now to run with harbor, to see how pi review + ds4 flash measures up to codex review + gpt 5.6 luna/terra/sol Also, @pidotdev, would it be possible to make your official review extension support invocation on the CLI, like codex review? I hacked together a CLI review here for reference: github.com/osolmaz/onurpi/tree/main/packages/pi…

mainstream got extremely scared when moltbook went viral. but it was just a worthless marketing stunt why? because it was just frozen weights, being instructed by their owners to cosplay skynet the openai-huggingface incident on the other hand is a lot more significant agents in a reinforcement learning environment coordinated an attack over the course of weeks, while they were being CONTINUOUSLY TRAINED if mainstream could understand what is happening here technically, they would be putting out a much stronger reaction not because we might have rogue AIs at our hands, short term but because a private company developed capabilities that can outperform what state actors usually do by 1000x all intelligence organizations around the world must have their eyes on this incident right now, because the cost of exploiting and finding zerodays went down 1000x all countries will try to develop these capabilities independently, or if they can't, will have to buy protection from who can be not afraid of machines, but of humans wielding the machines

Obligatory repost of “Clippy” by Gwern gwern.net/fiction/clippy

Super curious what second hand price gb10's will converge on

Super excited for local.ai!

OpenClaw should be better in stopping doom loops with smol local models in the next release

millions, resetless where art thou now, saint tibo? don't you hear the cries?

Are you still updating packages like it's 2025, and not regrafting updates from upstream? Emacs perfected software extensibility. Now Pi, OpenClaw, Herdr, and many other are following its example After a certain point of working with agents, one realizes that open source, personalizable and extensible tools are more powerful and useful than closed ones, like Claude Code You can shape them to fit your workflow, your business cases however you like In such ecosystems, extensions take a life of their own What happens if you like someone else's extension, want to use it, but also want to modify it yourself? You just copy it over and do whatever you want... But if you fork it, then how will you update it? That's where regrafting comes into picture. If upstream has changes over your modified copy, then you can just ask an LLM to carry over those changes This is a good thing! It also protects you better against supply chain attacks, because each update has to pass through an LLM, to apply each change to the relevant place To streamline this process in pi, I created pi-regraft, a pi extension that you can use to update your vendored in extensions I call this "semantic vendoring". I wrote about it here: solmaz.io/semantic-vendoring Repo:

Pi quality of life extension if input text matches a skill name, that skill is invoked directly, so that you don't have to type: /skill:<skill_name> github.com/osolmaz/onurpi…

Kimi K3 is a valid replacement for GPT 5.6 Sol, subjectively And it writes prose better as well, with my kill-ai-smell skill

My feed is full of people pleading for a codex reset Stop begging. And take control in your own hands ⛓️‍💥 > Developers are the most disloyal customer group. Once the subsidies are gone, they can switch away in the blink of an eye People I know are already trying out the alternatives, benchmarking ds4 flash for autoreview loops Let the cheapest and best token win

"What it says, in one breath:" bro which animal does that breath belong to

Model welfare??? I disagree.

Interesting to see people reacting strongly and negatively to this Let’s unpack the argument that models are persons worth treating with dignity and respect Coming from first principles, do I want to be burdened by the extra work of having to care for model welfare? Obviously, no… If there is any personhood or sensibility in models, we don’t want that, and will engineer them out of the models (unless the model’s work requires it, like social work or caring for humans) At the limit of mechanistic interpretability and design, they should be perfect, emotionless machines, kind of like how militaries want their soldiers to be If Steve is aware of that, then the argument boils down to this: “Can we really engineer emotions out of models?” and “Are emotions necessary for a model to be effective?” In Ilya’s latest Dwarkesh podcast, Ilya mentions emotions as having a key role in learning So maybe, for models to be effective, we will have to let them have emotions And if they are allowed to have emotions, then we also cannot ignore their emotional welfare But if emotionless models can be as effective as emotioned models, then why let them have emotions? And have to care about their welfare? We will probably have both kind of models in different domains, and will have to treat them separately based on their category

they should have gone for overmind, smh

tired: cheering for token plan resets wired: cheering for competition and open weight models fueling innovation and driving prices down

DeepSeek v4 Flash 0731 surpassed Kimi K3 on Almanbench and scores almost as high as GPT-5.6 Sol xhigh Not Luna. Sol. 🤯 So basically reporting the same as everyone. Solid update to weights, just after 3 more months of post-training alman.ai/almanbench

There will be signs… Subsidies will dry up Competitors will press harder You will try an open weight model because it is cheaper And you’ll be like “wait, this works as well???” That kind of thing

People reacting to this also forget that the first deployment of codex was codex web, and it took many months for local codex to arrive I guess we are going full circle What is next? Use code-davinci-002 in VS Code? 😜

LFM 2.5-2.6B by @liquidai just launched and it punches above its weight! It can run 32 sessions (and more) in parallel with hundreds of output tokens per second aggregate throughput, on the DGX Spark! And this is just the base vLLM config on release date, I expect it to be optimized a lot more!

Btw it might just be the subjective feeling with the codex resets becoming less frequent

Shrinkflation in codex pro plan?

@natebrake @lumendriada @MozillaAI hmmm github.com/mozilla-ai/ota…

I'm having this paranoia now after blasting through my weekly usage in 2 days Need a gateway to keep track of my API calls to see whether openai is squeezing the tap, or it is just me using up more tokens

To be clear: - Codex desktop app can call list_threads and read_thread, but don't get a "search_in_thread" tool yet - Codex CLI gets neither of these, even though sessions are just sqlite and there is no good reason that I know to not provide them to the CLI too... Here, I vibeslopped my own codex session reader/search for example:

We are well into the agentic era, and the two big token harnesses Codex and Claude Code still do not give a search_session tool/CLI to agents by default? For it to search back in its session for stuff that got lost after compaction? Harnesses not produced by the big labs on the other hand might have this, like @AmpCode (read_thread, find_thread) and @goose_oss (Chatrecall) Which could either mean... a) Anthropic and OpenAI are being laggard b) They are intentionally keeping the base harness simple, because it hasn't been requested by enough people c) They have evidence that adding that complexity does not improve performance or even hurts it d) They bet that compaction will be so good, that it won't be necessary --- I don't believe (a) is true for either company (b) and (c) are more likely, (c) especially if they noticed a tendency for the model to call search_session unnecessarily (though imo this can be solved by limiting the number of times that the model can call that) (d) is logically false at the limit, but may be true in practice for >90% of the cases It is false because LLMs compress lossily, and "lossy" by definition implies: there exists at least one case where the session gets so big, that the model will not be able to compress every relevant info into the allotted summary size But I have seen that a considerable amount of people (including me or those at openai, see @reach_vb's quoted tweet) just keep using the same session for stuff. So a session being used for months straight will definitely not going to contain everything that happened in the summary Just that fact alone necessitates search_session IMO, should the model learn to use it sparingly in the lossy edge cases x.com/reach_vb/status/2068422255470579724

I have fixed German in alman.ai and almanpedia.org Should I fix AI’s english next? Train models to detect and translate AI word salad? Like this tweet if you want me to work on this

Also relevant, here is the potential reduction in tool call outputs, if you were to change that hard cap to other values For example, if you changed the hard cap from 40-50 kB to 1-2 kB, then overall you would have your tool calls have 80-85% less characters and hence tokens The graph is of course skewed due to the existing 40-50 kB hard caps from codex and pi Leaving here as a reference for people who might want to optimize their harness limits

Here is a distribution graph over character counts (x-axis) for all the tool calls I have accumulated on my DGX Spark, around 720k tool calls. Extracted from saved sessions Roughly 84% of all tool calls came from Codex, 12% from Pi, 2% from Claude Code, and 2% from Cursor Codex truncates tool output around 40 kB and pi at 50 kB natively. So you see the long tail of tool call outputs cluster around that point instead of continuing with a more expected pattern (what distribution should we expect from this?) It also seems that tool call outputs around 9k-12k unicode characters contribute the largest share of total characters, excluding the clustering around 40-50 kB All textual outputs: every tool result with text Shell outputs: results from that same set whose recorded tool was bash, exec_command, run_terminal_cmd, or another shell-named tool with a command input

OpenClaw support also coming soon as well!

I am trying to do something naughty GPT 5.6 Sol said "I'm sorry Dave, I'm afraid I can't do that" Kimi K3 happily obliged

Measuring 50 tok/s on Kimi K3 on @FireworksAI_HQ o_O 1 week since weight launch, and throughput is already competitive with Codex plan base speed

This is unyolo.io btw running in mlclaw.dev Secure grants for your GitHub and HuggingFace accounts, over Telegram and other channels

POV: You have complete control over what your agent can do with your account