And in my optimistic view, an aligned superintelligence is more likely to work towards collective benefit rather than blindly listen to its (lower intelligence) creators.
Feels like we're turning scientific discovery into a race of: will my Claude find it first before yours?
Like any other race dynamic in history, humanity probably benefits as a whole. But it is nevertheless sobering to see.
Hopefully future AI systems work on problems that humans collectively decide are important, rather than concentrate power into a few (unaccountable) hands.
is the Anthropic IPO chart a better LR schedule than warmup cosine decay?
let's find out! @mridul__agarwal@ananthsiyer and i are running 4 identical pretraining runs, with the LR driven live by ANT, SPCX, SPY, and NVDA.
watch the models train at https://t.co/z9JTXXLpkg
@yminsky Love it! We’re doing similar work @UT.
A small sneak peek, a more substantial piece coming out in a few weeks :) Formal verification is the future of programming, no human is capable of reading the volume (of slop) code
https://t.co/kJHy6BQ4Eq
Can a formally verified C compiler be as fast as gcc?
CompCert’s optimizations come with a machine-checked proof that they don’t change your program’s semantics. The catch? The compiled code runs ~28% slower than gcc -O2.
We tried closing that gap with AI agents 🧵
Honestly, all these problems exist with GPT 6 Astra as well. Haven’t tried 6.2, but wouldn’t be surprised. The Sol models have been much more “usable” in my experience. But no GPT model has come close to Opus 5.5.
@ariccio Agree! With AI Agents generating code much much faster than humans can verify them, formal methods can give us a smaller trusted boundary to review and allow us to trust the spaghetti that is generated :)
Can a formally verified C compiler be as fast as gcc?
CompCert’s optimizations come with a machine-checked proof that they don’t change your program’s semantics. The catch? The compiled code runs ~28% slower than gcc -O2.
We tried closing that gap with AI agents 🧵
@P33RL3SS I believe this is exactly what we are asking: CompCert gives us correctness, but it is slow. 1) Why is it slow? 2) Can we make it faster while being correct?
Coding agents made passes and proofs cheap. The hard part now is verifying the model and assumptions the proofs rely on, and making sure they match reality.
Full write-up: https://t.co/CEHU9qkIKh
Where we landed, vs gcc -O2:
- CompCert benchmarks: 28% behind → 4–6%
- PolyBench loop nests: 55% behind → ~14%
- Held-out larger programs: 13% behind → ~9%
On the first two, we’re now faster than gcc -O1.
@thsottiaux Codex Desktop is very slow, lag inducing, clunky, and consumes too much memory. The difference is especially stark after using Claude for a while.
I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model.
Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59. It also leads across key cybersecurity evaluations scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching.
We’re seeing significant improvements in long-horizon tasks even compared to Argon, and we’ve strengthened frontier safeguards.
As we did with Argon, we’re going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week. We'll be expanding it to more partners through our Fairwind Program, ahead of wider availability soon, so hang tight!