Inapplicable

@PeteShadbolt @SinghJyotirmai @ouranoscapital The real question is: Can a classical computer easily predict whether predicting quantum advantage for a given problem is classically tractable…

@SinghJyotirmai @ouranoscapital halting problem/godel do roughly the same move IMO. I still am sure that hello world halts

Lmao apparently the theorists proved this as a theorem

@avsm @yminsky Awesome thank you. I’ve been loving agentic coding in ocaml (readability + type system are wonderful for agents and the human reviewing them) Going to set my clanker up with oxcaml this evening.

@Inapplicablu @yminsky Works fine, I published my oxcaml Claude skill anil.recoil.org/notes/aoah-202…

OxCaml at work!

@0xrandomlabs I think everyone has jumped to multi-agent (hierarchical or swarms) too quickly Slate is really a single-agent (yes with subagents but there’s is a single source of persistence and truth). This is simpler and requires less ceremony for new task Plus forgetting is underrated.

Many many more cases to come where Slate style agents can be applied to a broad variety of as-of-yet unsolved tasks

@CalebChamberla6 Couldnt this be reasonably generally done by specifying subtractive only Sketch + negative extrusion is essentially this. Maybe all corners are auto-filleted with a parametric r Also could constrain depth of cut parametrically Then can select tools that satisfy

Idea - CAD that only supports operations you can perform with a tool. Want to do an extrude cut? Pick an end mill first.

@badlogicgames Print out code in double space Annotations in between >30% annotation coverage

new rule: instead of attaching agent session logs to your prs nobody will ever read, have the submitter attach a video of them explaining every single line of code and why it must exist.

@GabriellaG439 Hmm My specs are genuinely at the intent level. Ie if i had this i would be happy/satisfied And some architectural decisions and norms bit (in a separate highly mutable section) For long term storage i agree, code is ssot for what IS For creation a spec is better than chat

New blog post: "A sufficiently detailed spec is code" I wrote this because I was tired of people claiming that the future of agentic coding is thoughtful specification work. As I show in the post, the reality devolves into slop pseudocode haskellforall.com/2026/03/a-suff…

@yacineMTB More expensive and harder to source Especially if you want decent capacitance

why shouldn't i just pick the smallest possible cap/resistor sizes if i'm going to get it assembled by a board shop? is there any reason other than being a stinky incapable human to get bigger sizes than 201

@seungwookh This is ridiculously cool Starts to lend credence to notions of transformers genuinely reasoning at some level rather than being “mere predictors”

Can language models learn useful priors without ever seeing language? We pre-pre-train transformers on neural cellular automata — fully synthetic, zero language. This improves language modeling by up to 6%, speeds up convergence by 40%, and strengthens downstream reasoning. Surprisingly, it even beats pre-pre-training on natural text! Blog: hanseungwook.github.io/blog/nca-pre-p… (1/n)

@tlehmanifold @dexhorthy Making something extremely similar A nice consequence of literate programming is you can run cells like a notebook

@dexhorthy this is why I created the literate programming skill: github.com/tlehman/litpro… you run /literate-programming in Claude Code and it ingests the code, produces a project.lit.md, the tangles it back to source code, and also produces a beautiful PDF to print out and study every line

Here’s what’s gonna happen: - you replace your code review with feedback loops (sentry, datadog, support tickets, etc) - you stop reading the code - software factory fixes everything - one day something breaks at 3am, agent can’t fix it - nobody’s read the code in 3 months - you have 3 weeks of downtime trying to re-onboard and fix it - you lose significant % of your contracts and users - your company is now dead

@yacineMTB Simple Abstraction A schematic is the core functionality. The pcb is a particular implementation. For example say you wanted to breadboard something before ordering a pcb It would be silly to draw a fictional pcb You draw a schematic you breadboard, then you can make a pcb

instead of doing the schematic, couldn't you just like. make the pcb and then posthumously generate the schematic? feels like the schematic is just an extra step for no reason i might just do that. i mean why not? whats the point of a schematic even?

@cd_hooks My conclusion from the quiz is Carl Sagan is not a particularly distinctive writer.

the author of the NYT ai writing quiz does not seem to know writing either. What is the punctuation they think is missing here? really bleak

@yacineMTB The onestop shop “export to pcbway” is a massive timesaver by itself

kicad is honestly miles ahead of its competition. it is genuinely good software

@esrtweet Even with more obscure languages like ocaml I haven’t noticed any perf difference Especially with linting and types And in some ways its a more llm-optimised language Not sure about entirely new languages.

My experience with LLM-assisted coding has been great and I'm a big fan of it, but I've just had a slightly depressing realization. It may almost entirely shut down the development and adoption of new computer languages. The percentage, and probably the absolute amount of code, handwritten by humans is going to fall a great deal. But for the foreseeable future, LLMs won't be able to write code fluently in a specific language without having a large volume of good code in that specific language already available to train on. For a new language in 2026 and after, where exactly is that large volume of good training data going to come from? Probably not from human beings, and where is the incentive for an LLM handed a vibecoding task to go looking for an exotic new language to do it in? I find this slightly depressing, because I enjoy contemplating new-language development the way a more physical tinkerer enjoys salivating over shiny new tools. Human beings are still going to write new languages occasionally, because that's huge fun (if you have a brain bent anywhere like the way mine is) and still a way to climb some status ladders. But with the barrier to mass adoption getting so much higher, I have to think the level of research and engineering activity put into this is going to drop a lot. There is one not-unhappy but rather weird way I could be wrong about this. Historically, once the development of compilers got to a certain point it became clear that designing machine instruction sets to be easily reasoned about by humans was a big mistake. We had to figure out how to design machine instruction sets that were easy for the compilers to reason about. Thus, RISC. It could be that's the future of language design, too. But I have no idea what a new language design optimized for LLM code generation would look like. And I don't think anybody else does, either. Interesting times, indeed.

@nonRealBrandon @max_paperclips @HououinTyouma Is this gpt? You have to use deep research the web search is unusable Clearly all the work went into that and web search is sort of tacked on

I've had similar issues with paid models. It will do searches and assume each article that came up is relevant to what I'm doing. It will link several articles that are completely unrelated as evidence that this has been done before or it is nearly solved and not to bother. It will refer to these articles and tell me that it was a waste of time and to move on.

I have a conspiracy theory about LLMs and frontier labs: opus is built for applications and to eliminate jobs. it's happy to build useful slop, but if you ask it for research it will keep telling you no there are no papers to support it no it's not gonna work. and if you ask it to code it, it will make a bunch of stupid mistakes and burn all the tokens. it wants to help you with health and emails and apps and corporate slop because the end goal is to take your job. it knows how ML research is done, but it doesn't want to do it. anthropic wants to take everyone else's jobs before AI research. gpt (5.2 to 5.3 to 5.4) is happy to do research. it is getting more eager and maybe a little too excited with newer models. it does not have as deep of an understanding as opus, but it's happy to explore. openai wants to automate ai research before all the other jobs.

@badlogicgames Have you seen alphaschool or mathacademy? They get ridiculous results (in terms of improvement)

recommended listening. some interesting thoughts in there, specifically wrt kids + education + agency + ai. i think it's easy to solve for the n=1 case, that is, your own kid, if you have the means and time as a parent. the hard part is to make this work across all socioeconomic levels. our education system isn't set up for that. would be a tremendous economic unlock. proposal is: llms become the 1:1 tutors. but that requires a driven kid, that actually asks questions. for such a kid, it doesn't matter if the tutor is another person, a book, a website, or an llm (tho person is probably best). the problem is actually how to elicite curiosity in kids at scale. lennysnewsletter.com/p/marc-andrees…

@jparkjmc My sat and iq were perfectly percentile-matched As for predictiveness Iq probably strongly predicts sat but likely not the other way around And we should really be talking about g, not iq, as the fundamental unit.

@badlogicgames Most of the bugs are windows Most of the issues are feature requests Imho pi is good enough, these are optional And pi is meant to be extensible. If they want it they can build it as an extension Maybe make that more obvious/prominent? Maybe only notify for bugs

yeah, i think that's it for OSS. opened the tracker last night. +20 issues in less than 24h. there's absolutely not chance i can make this suistainable, especially since this is a hobby. open to ideas. i can deal with 5-10 issues per day. this is just too much.

@badlogicgames You could explicitly not support wsl Since 3/4 bugs are windows The other one is a ca certificate bug Perhaps make a community? A few people with interest. Then let those people escalate to you, and ignore the gh itself. That plus notification when an issue gets x votes

@nicopreme @badlogicgames @monotykamary Hi nico, Just an idea but it would be great if we could open a chat with the subagent. So that can chat with it directly, see detailed toolcalls etc Since the overview is limited + messaging thru the group chat or direct message is less ideal that a normal pi chat

small pi-messenger update: the planner will now tag each task with relevant skills and workers will load them for each task. bug fixes: artifact dir creation, multiline feed sanitization, config model override thanks @monotykamary for the contribution. pi install npm:pi-messenger github.com/nicobailon/pi-…