XC (Xiaohong Chen)

CS PhD @UofIllinois | math, logic, code, AI

[Nathan Labenz @labenz]

@RosuGrigore on AI:AM, on safe AI: it's "impossible to prove the weights correct."

But you can write safety specs — this should not happen — and prove the monitor that keeps the system inside the safety envelope. A police system around the AI.

https://t.co/oidAERHAmc

[Harsha Gundala @harshagundal]

They were building in stealth for 2 years, I was building in stealth for 2 hours…

Happy to open source Qwen-2.5-1B-RLCD, 5x faster on-device inference for JSON workloads that need to be type-safe.

⚡️Demo below on a M4 MacBook⚡️

every LLM has the ability to efficiently batch inference every key of a JSON at the same time and generate probabilities from a set of possible categories. No new training required, but it’s easy to optimize if you need!

On hugging face now!

[Grigore Rosu @RosuGrigore]

1/4

Program execution IS proof generation. This becomes literal within semantic frameworks, where program execution is rooted into formal language semantics.

The same foundation gives us another opportunity: training data generated directly from code.

When execution follows formal language rules, every step has a justification. Compose those steps, and a finite execution becomes a proof of how the program moves from its initial state to its final state.

Across unbounded programs and inputs, code supplies infinitely many finite execution proofs. We can retain the program states, the rules applied, and the derivation connecting them.

[Grigore Rosu @RosuGrigore]

1/6
The next scaling law for coding models may be intelligence-per-token, or ITP: how much reliable reasoning each token buys.

Benchmark scores keep climbing, while practical context windows seem to move much more slowly. Context will grow, but it will always be finite and expensive.

At a fixed context budget, a model needs to do more with the tokens it has. Better representations and more efficient reasoning could let it solve problems that would otherwise exhaust that budget.

[Fast @fastxyz]

A festival floor is the hardest stress test in consumer payments.

Last month at UNTOLD, Fast settled 160,589 operations onchain, through our joint pilot program with @RhunaIO.

While most networks take minutes to reach finality, card settlement takes days, Fast reached finality in ~3 ms.

[XC (Xiaohong Chen) @cxh93]

@RosuGrigore I used https://t.co/cG8A5MDmWz for this purpose!

[Grigore Rosu @RosuGrigore]

If you work on a hard problem and keep partial results private until the final proof, you may worry that the AI model you use, or the company behind it, ends up holding your ideas. Secrecy alone will not fix that, because you have to type the ideas somewhere. What you can do is make the date provable. Then, if your work later appears inside something larger, you can show exactly what you had and when you had it.

This is free and takes minutes. Put everything you want to claim into one archive: the paper, the notes, the code, the git commit id, and a short plain-text abstract saying what the snapshot is. Take the SHA-256 of that archive. Timestamp the hash on the Bitcoin blockchain with OpenTimestamps, or on your favorite decentralized network. Only the hash leaves your machine, so nothing is disclosed. You then get a small proof file that ties your hash to a specific block, self-contained, with no trusted third party.

1. Bundle, hash, stamp. Keep the archive and the proof file together, offline.
2. Your coding agent can do all of it. Tell it to build the bundle and run the stamp. Your only job is to keep the archive safe.

It does not prove authorship and it does not stop anyone from using your ideas. It proves that those exact bytes existed before a given block, which is enough to say later, with evidence, that the work was yours on that date.

I do this now with all my projects, at every milestone.

[XC (Xiaohong Chen) @cxh93]

And here are the links to the public repos; please request if you wanna access to the private repos for more details:
📊 SWE-bench audit, all 86 rewrites with per-instance analyses: https://t.co/V9Kme1eIXJ
📚 Korpus, first public sample: https://t.co/v1X5mHGaD1
✅ Program verification proofs: https://t.co/2RNC2BOJkC
🐞 Real-world defects case studies: https://t.co/vHLYaIdwcs

[Grigore Rosu @RosuGrigore]

Coding models are running out of data 🧵

The ability of LLMs to code depends heavily on the data they see about code, and lately the industry seems to have consumed essentially all high-quality code data for training and evaluation. Progress is visibly slowing. Whoever finds the next big pool of data holds the key to a new door.

My group came from programming languages and formal methods, and we know where such a pool is. More than that, we believe this data changes what an AI answer is: code should ship with its specification and a machine-checked proof, so we no longer need to spend the human time to review the code itself.

We call this new paradigm intent computing, and we ran several experiments...

[XC (Xiaohong Chen) @cxh93]

@ChrSzegedy If AI is able to solve IMO questions in minutes then it should have no problem writing pre/post conditions and finding the right loop invariants, and do end-to-end formal verification while it codes.

[XC (Xiaohong Chen) @cxh93]

@ChrSzegedy Strongly agree. Isn't there a lot of recent interest in training on Lean proofs? And with formal methods that can be extended to general code & computation.

[XC (Xiaohong Chen) @cxh93]

@Sergey007S @fastxyz @RosuGrigore Please check more details from our docs: https://t.co/sARwN3GgOC. Look for external claims transactions.

[XC (Xiaohong Chen) @cxh93]

@Sergey007S @fastxyz @RosuGrigore @Sergey007S And, the overall message can be checked against the network's validators.

[XC (Xiaohong Chen) @cxh93]

@Sergey007S @fastxyz @RosuGrigore Hey @Sergey007S , thanks for the message! Yes, an external claim settled on Fast is independently checkable. The verifiers' signatures can be checked against the public addresses of the verifiers' addresses and the content of the tx.

[XC (Xiaohong Chen) @cxh93]

https://t.co/fjbSj0GwkP

[XC (Xiaohong Chen) @cxh93]

I've started to use normalpowers (link in reply) to guide my LLM model to think and brainstorm, and it's quite convenient and useful. I feel that it makes AI models more structured when they think, which is nice, because otherwise they tend to simply "infer" the output and conclusion solely based on the tendency of the context and the input.

Basically, it helps to get better prompts for LLMs without me spending a lot of time on it. Worth a try.

[Grigore Rosu @RosuGrigore]

https://t.co/0Iu03ZZCFm

[XC (Xiaohong Chen) @cxh93]

@Sergey007S @fastxyz Good question! Verifiers do "off-network" checks that client wants to do. Verifiers are regular accounts who can sign a client's signed message, and validators check if sufficient signatures are collected. Quorum = 0 simply means that no verifier signature is required.

[XC (Xiaohong Chen) @cxh93]

Good tool + Bad skill = Bad UX. Optimizing skills over repeated benchmarks is a much better path than prompt folklore indeed.

[Fast @fastxyz]

skill-optimizer open-source repository:
https://t.co/nQWJ5uz2IV

[Fast @fastxyz]

Your AI agent can now buy things on Amazon while you're on the go.

No card. No login. No Amazon account needed.

Fund with USDC → tell it what you want → it orders it and sends you a tracking link