jsmorph

@epistematics_ai

[jsmorph @jsmorph]

@Altimor https://t.co/kbvcxbtIlo

[jsmorph @jsmorph]

@DimitrisPapail As an experiment, I've banned words and phrases like "load bearing", "smoke test", "wire in", "landed", "shape", and "guardrails" in favor of specific, clear, and plain alternatives. I think I see the LLMs generating better code and technical work as a result. Dunno.

[jsmorph @jsmorph]

@ClaudeDevs Easy and maybe better with tuple spaces (1985). https://t.co/lKbzQPkwZ1 https://t.co/ZKzI9TBiYz

[jsmorph @jsmorph]

@0xz80 https://t.co/mpfzbJnwYM is starting to verify emitted WASM.

[jsmorph @jsmorph]

@0xz80 https://t.co/jcQpxCbSTf

[jsmorph @jsmorph]

@ylecun @gabriberton Here's some very impure but verified code generation: https://t.co/AVdVt36bou

[jsmorph @jsmorph]

@UnderwaterBepis @lanyon_ai That's what I was asking about (presumably including formalization of IEEE-754, a subset of C, etc).

[jsmorph @jsmorph]

@JunoRavin @lanyon_ai @getjonwithit @minmodammar Following up to myself to offer this link, which provides a lot more background. https://t.co/Oj0W9j2ojU

[jsmorph @jsmorph]

@JunoRavin @lanyon_ai @getjonwithit @minmodammar Thanks! I was looking at that benchmarking post. More details about that C evaluation would be great. Same with that IEEE-754 work. The symbolic work on the DSL is cool.

[jsmorph @jsmorph]

@UnderwaterBepis @lanyon_ai I think the C code is generated from the DSL, not Lean. Not certain. But that doesn't suffice for verification.

[jsmorph @jsmorph]

@elonmusk Maybe even an efficient and formally verified binary https://t.co/6G5w7yOtQU

[jsmorph @jsmorph]

@lanyon_ai Don't get me wrong. Looks very cool.

[jsmorph @jsmorph]

I'm not really seeing the end-to-end proofs of correctness through to C. (I do love the Racket DSL!) If the C analysis is based on sampling, how is the sampling performed? How are the criteria established? I guess the double vs Real trouble is hard to avoid. 1e-8 and all that. C semantics itself I guess is way out of scope.

[jsmorph @jsmorph]

@TrisH0x2A @jartine 's Lisp is very cool. https://t.co/sxu27CLTst

[jsmorph @jsmorph]

@littmath @ElliotGlazer

[jsmorph @jsmorph]

I want verifiable extraction to a back end that itself is verifiable -- or at least somewhat so. That's why I was trying WASM with https://t.co/elwdytwtYB. My compiler is in Lean of course. It's hard to prove that the generated WASM is correct, but it is sometimes feasible. Here's an (undocumented, sorry) proof that some WASM really does correctly implement a central limit order book for limit orders. https://t.co/9Y7bLrGpop

[jsmorph @jsmorph]

@OpenAI SheerID doesn't work.

[Elliot Glazer @ElliotGlazer]

Leo has updated the Lean kernel to fix this bug though it appears Comparator itself is still vulnerable to this. As always, more thorough discussion can be found on the zulip: https://t.co/hiAwRIndjC

[jsmorph @jsmorph]

@getjonwithit How is the C code formally verified?

[Alex Kontorovich @AlexKontorovich]

I would add just one more stage to this: we need to verify that the Lean *statements* and *definitions* actually say what the natural language proof says. This “semantic alignment” problem is not solvable in silico! I’ll mention this in my talk on Monday night. #ICM2026