Entries for May 2026
-
-
This looks very promising! -
-
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 -
i0 to i4 interest scale
One cool small invention in engineering management is the
p0,p1,p2,p3,p4priority scale.It compresses a lot of social and operational context into two characters. Lower number means higher priority. More importantly, priority is tied to action. If something is
p0, somebody needs to do something.But there is another scale I want for personal knowledge work:
i0,i1,i2,i3,i4.The
istands for interest. Priority is for actions. Interest is for attention.- If
p0means “act now”,i0means “do not lose this”. - If
p1means “schedule work”,i1means “read soon”. - If
p2means “do later”,i2means “useful context”. - If
p3means “low priority work”,i3means “weak signal”. - If
p4means “almost never work”,i4means “almost never revisit”.
This is useful when you need to rank interest concisely across many topics, sources, or articles.
For example, you might follow several sources about the same broad topic. One source is must-read, another is useful background, and another is only worth keeping around for occasional context. They are all about the same thing, but they do not deserve the same amount of attention.
I use this for myself in Scoop, a news intelligence system I am building to collect articles, group related ones, and rank how much attention they deserve.
- If
-
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 -
And what good is this for? It lets me program my claw @dutifulbob to extract signal from the noise, and display it in my personal open source news aggregator scoop I feed it discord messages, openclaw git history, and other various sources, and it's supposed to evaluate whether that content deserves my interest. it's still work in progress, because the more batched you process all the info, the worse it informs in the screenshot below, my claw underrepresented what Peter has done in one day 👎 on the other hand, it has also found a PR about local model discoverability 💪 Here is the system I use to aggregate all my info, still under development: github.com/dutifuldev/scoopImage 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! -
Excited to work with @steipete, @vincent_koc, @LysandreJik, @ben_burtenshaw, @evalstate, @mervenoyann, @NielsRogge and many others! -
-
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