mutkach

True, and even selling token price may not be illustrative of what it actually costs them to provide the service. Tokens may be sold at a loss if majority of spenders are running agents 24/7

The industry and users moves on from single chat-based to more and more "agentic" workflows that may generate longer workloads with multiple simultaneous agents (separate agents - separate contexts - separate KV caches).

My estimation is based on, say, running a Kimi3 on a 24 B200 GPUs - it is very easy to lose money when selling tokens at "market" prices.

One thing that I tried was porting an old and obscure 32-bit VST plugin from PowerPC to modern architectures

Any recommendations for non-typescript-based coding TUI harnesses/agents other than Codex? Given that Codex is open source someone should've built an untethered client?

Please consider adding `cargo watch` - that would be a killer feature!

/*

* Check if 1M context is disabled via environment variable.

* Used by C4E admins to disable 1M context for HIPAA compliance.

*/ export function is1mContextDisabled(): boolean {

  return 

}

Interesting, how is that relevant to HIPAA compliance?

Super-relatable.

Now that I think about it, most of my advice starts something like "Here's what you're gonna do..."

Wait, that itself sounds like a problem, but how do I fix it...

You are correct, that was an uneducated guess on my part.

I just glanced at the IR which was different for some attributes (nounwind vs mustprogress norecurse), but the resulting assembly is 100% identical for every optimization level.

Probably LLVM runs different sets of optimization passes for C and C++. Need to look at the IR, or assembly to know exactly what happens.

Sure, he is one of biggest advocates for it, and yet he was quite clear that it is not yet possible for him to do his actual research in Lean.

Quoting one of the recent papers (2020):

With current technology, it would take many person-decades to formalise Scholze’s results. Indeed, even stating Scholze’s theorems would be an achievement. Before that, one has of course to formalise the definition of a perfectoid space, and this is what we have done, using the Lean theorem prover.

I share your fascination with proof assistants and formal verification, but the reality is that I am yet to see an actual mathematician working on frontier research who is excited about formalizing their ideas, or enthusiastic about putting in the actual (additional) work to build the formalization prerequisites to even begin defining the theorem's statement in that (formal) language.

Understanding IMO is "developing a correct mental model of a concept". Some heuristics of correctness:

Feynman: "What I cannot build. I do not understand"

Einstein: "If you can't explain it to a six year old, you don't understand it yourself"

Of course none of this changes anything around the machine generated proofs. The point of the proof is to communicate ideas; formalization and verification is simply a certificate showing that those ideas are worth checking out.

more than 1,000,000 lines of Lean 4 code and concluding with a QED.

Usually the point of the proof is not to figure out whether a particular statement is true (which may be of little interest by itself, see Collatz conjecture), but to develop some good ideas _while_ proving that statement. So there's not much value in verified 1mil lines of Lean by itself. You'd want to study the (Lean) proof hoping to find some kind of new math invented in it or a particular trick worth noticing.

LLM may first develop a proof in natural language, then prove its correctness while autoformalizing it in Lean. Maybe it will be worth something in that case.

Before Church there was Peano, and before Peano there was Grassmann

It is rather well-known, through Peano's own acknowledgement, that Peano […] made extensive use of Grassmann's work in his development of the axioms. It is not so well-known that Grassmann had essentially the characterization of the set of all integers, now customary in texts of modern algebra, that it forms an ordered integral domain in which each set of positive elements has a least member. […] [Grassmann's book] was probably the first serious and rather successful attempt to put numbers on a more or less axiomatic basis.

What would you suggest as a reference problem (a benchmark of sorts) to try to play with formal methods for someone with just a bit of formal verification background but not in the field of software verification? Can you suggest some helpful materials?

I've come across TLA+ multiple times, but it seems it was more targeted towards distributed systems (Lamport being the creator, that makes sense). Is it correct, that it would be useless in other domains?

What do you mean exactly by "error-tolerance"? Is it like, each node is wrapped into a result type, that you have to match against each time you visit it, even though you know for a fact, that it is not empty or something like that?

I suppose that one of the pros of using tree-sitter is its portability? For example, I could define my grammar to both parse my code and to do proper syntax highlighting in the browser with the same library and same grammar? Is that correct? Also it is used in neovim extensively to define syntax for a languages? Otherwise it would have taken to slightly modify the grammar.

I certainly hope so.

I wonder, what is the actual blocker right now? I'd assume that LLMs are still not very good with specifications and verifcation languages? Anyone tried Datalog, TLA+, etc. with LLMs? I suppose that Gemini was specifically trained on Lean. Or at least some IMO-finetuned fork of it. Anyhow, there's probably a large Lean dataset collected somewhere in Deepmind servers, but that's not certification applicable necessarily, I think?

AI also creates a need to formally verify more software: rather than having humans review AI-generated code, I’d much rather have the AI prove to me that the code it has generated is correct.

At RL stage LLMs could game the training*, proving easier invariants then actually expected (the proof is correct and possibly short - means positive reward). It would take additional care it to set it up right.

* I mean, if you set it to generate a code AND a proof to it.

LLVM makes it so much easier to build a compiler - it's not even funny. Whenever I use it, I feel like I'm just arranging some rocks on a top of a pyramid.

What language do you use parser combinators in, and what kind of grammar do you parse usually? Nom was terribly verbose and unergonomic even by Rust's standards. Haskell's Megaparsec/Parsec is good but yeah, it's Haskell, you need to handle multiple monads (Parser itself is monadic, then your AST state, and maybe some error handling) at once and that's where I got confused. But I appreciated the elegance.

I experimented with PCs in Haskell and Rust (nom), then moved on to parser generators in Rust (pest.rs), Ocaml (Menhir), Haskell (Happy) and finally ended up with python's Lark - the speed of experimenting with different syntax/grammars is just insane.

Does it support sending and executing commands to the panes like tmux does?

like this:

tmux send-keys -t 0:1.1 "ls" Enter

edit: well, yes, you can:

zellij action write-chars ls

zellij action write 10

That "itch" is exactly what I meant, lol! And I agree! I'd definitely give Rust a try. Playing around with types and traits until they click is genuinely addictive - it feels like solving a puzzle or something

More information is needed to give proper advice:

- Do you like filling out the type annotations in Python (making sure linter check passes)? Do you like TYPES in general?

- Do you like working with memory (crushing memory bugs, solving leaks)?

- Do you prefer imperative or functional approach more?

I remember Arc randomly rewriting my bookmarks using some kind of summarization model or something like that, it also sometimes changed the name of downloaded files reinterpreting their names. Maybe it is related somehow.

Well, I guess it was the first AI-first browser, hence all this bs. I uninstalled it months ago...

Another salesperson trying to jump on a departing bandwagon, without having an interesting product and not providing any valuable insights.

grief

denial

Yeah, and he's in the denial phase

Why this weird json output, though? Well, I figured it out - had to strip the newlines and some backslashes - but just having a few plaintext objects or files makes more sense, I think.

They seem to be betting on multiple things happening at once - as indicated by recent projections by their CFO, following Oracle's own insane projections. They are betting on demand rising exponentially and somehow keeping up with it using a consistent monetization model - which is not there yet. I am not a finance person, but I wish more people with qualifications - and hopefully unaffiliated with any party - scrutinized all those claims and projections.

1/6th of total US economic growth over the last year

What economy indicator exactly are you referring to?

How would you peer-review something like that? Or rather, how would you even _reproduce_ any of that? Some inception-based model with some synthetic data? Where is the data? I'm sure the paper was written in good faith, but all it leads to is just more crackpottery. If I had a penny for each time I've heard that LLMs are quantum in nature "because softmax is essentially a wave function collapse" and "because superposition," then I would have more than one penny!

Why would you really need something like that in a non-totalitarian state? Basically, it follows the russian playbook (essentially the same 'language' - safety concerns), but instead of the FSB, who is the beneficiary actor in this case?

55,000 hours of [IT] experience

I didn't know Mike Judge was such a polymath!

How would they not "share information with third parties". You need to sift through the data to make it even remotely useful for "training". You absolutely need to share it with either Amazon (for Mechanical Turk) or with Scale AI.

I am wondering how would you use a chat transcript for training? Unless it is massive, possibly private codebases that are constantly getting piped into Claude Code right now. In that case, that would make sense.