aminoacid

@kaepora is doing really awesome stuff :).

when you have a random name generator for your instances and it starts threatening you

@hthieblot confidential

pitch me your company in 1 word.

@michael_nielsen i think it also goes to show how much team chemistry matters in team sports.

Both their smiles! And the goal so beautifully illustrates both - Messi trusts Ronaldinho to do the impossible chip, and so he runs behind the defenders, rather than wide. It's a small matter, but illustrative

Leo Messi and Steph Curry remind me so much of each other in how they play - so much joy and creativity, so enabling of teammates, so loved by their teammates. It’s really beautiful to watch!

I'm in SF this week around HumanX. If you're thinking about private inference, secure model deployment, or confidential computing, I'd love to talk about what we are up to at @Confi_AI

Trusted Execution Environments, explained. A useful analogy is the transition from HTTP to HTTPS. With HTTP, you sent your data in plaintext but couldn't confirm who you were talking to. HTTPS allows you to confirm who you're talking to and encrypt your data in transit. But once your data arrives at the server, it's decrypted and processed in plaintext. This means the system administrator can see your data. TEEs go a step further. Your data stays encrypted by hardware during computation. You can verify who you're talking to, what software they're running, and that your data is private throughout. The sys admin can't access the data. No one can. All of this, like HTTPS, adds a negligible performance cost.

@realbarnakiss this is very interesting, are you planning to publish any source code for this?

I implemented zk-autoresearch, based on Karpathy's autoresearch loop, on a production ZK prover, Plonky3. Soundness review by a Plonky3 engineer is pending before I treat these as final. The methodology finding is already clear, preliminary results below. Target: Plonky3's NTT implementation — the inner loop of proof generation, already heavily optimized by expert ZK engineers. If the approach doesn't work here, it doesn't work anywhere. Hardware: Hetzner CCX33, AMD EPYC, AVX512, 8 cores. Model: I used Claude Sonnet 4.6 deliberately, Opus would have marginal gains at significantly higher cost per iteration. For a loop running potentially 100s of times in future experiments, that tradeoff matters. 74 iterations. Fully autonomous by design, but in this first experiment 2 adjustments were made to the setup (at iterations 5 & 10) to nudge the agents to be more decisive. - Raised MAX_TOKENS from 8192 to 20000, and added "you must always make a change" as the agent kept hitting the token limit. This unlocked improvements at iterations 6 and 9. - Added near-miss display in the history prompt, showing reverted experiments within 1.5% as combination candidates. This set up iteration 21, where the agent revisited a failed idea that now worked because the surrounding code changed. Iteration constraints: - Each iteration ran correctness tests to prevent faulty proofs. Note: during the run these were compile-level checks; post-run correctness was confirmed via full end-to-end ZK proof generation and verification with Radix2DitParallel on BabyBear (10 tests, all passing). - Agents were structurally prevented from touching FRI or other soundness-critical components — only dft/src/ and baby-bear/src/ were writable. 3% faster at the target size (2^20) during the experiment. Post-experiment benchmarks across 2^14 to 2^22 showed the optimizations generalized better than expected, particularly at the extremes (see image). The agent only optimized for 2^20. The known issues (short history window causing agent amnesia, wasted tokens on repo exploration, correctness test targeting wrong package) meant the last improvement was found at iteration 21. Round 2 with these fixed should yield a more consistent staircase pattern over 100 iterations. All gains came from the agent finding redundant work in the hot butterfly loop: precomputing products, hoisting broadcasts, skipping multiplications by 1. Pure implementation-level work, no algorithmic changes. 6 improvements in 74 iterations. 57 regressions. The full experiment log with every diff, benchmark result, and agent reasoning is auditable. The agent that found these improvements is not a ZK expert. It reasoned about Rust and Montgomery arithmetic from first principles and found real optimizations in code already written by expert engineers. ZK has been underexplored for agentic optimization because people worry about agents breaking proof soundness. The concern is real but misapplied here, all 6 changes are mathematically equivalent transformations, verified by end-to-end proof generation and verification. (Soundness review by a Plonky3 engineer is pending) Round 2 is being prepared with the known issues from Round 1 fixed. Full findings and code will be open sourced after it completes. If you are ZK team and want to run this, feel free to DM me. Inspired by @karpathy autoresearch pattern. First known application to a production ZK prover.

This is the type of insight that claude code has when i ask it to find patterns

We built a simulator for c8s, our confidential Kubernetes product with autoscaling.

claude-fleet now supports gcp and local dev containers cc @trq212 link below

generating so much output with claude code, the bottleneck has become the review process my a large margin

claude code is being extra dumb today.

@wyatt_benno @LouisThibault87 the CC functionality in GPUs is already available in H100s, B200s, B300s, and the new vera rubins. So they cost the same as the GPU itself, you just have to turn it on and pair the GPU with a CPU that also supports CC which is the case for most modern data center CPUs

@ameanasad @LouisThibault87 How much do those cost? Also what is the wait time to get one?

Perhaps I'm late to the party on this one, but seems like a more serious breach of TEEs than what I've seen so far. arxiv.org/abs/2602.11088

Progress :)

We just shipped Kettle, which builds and verifies *attested builds*. Kettle uses TEEs (Trusted Execution Environments) to sign builds using hardware attestation. Link in thead.

@gakonst @karpathy i started on something that runs claude code instances on confidential VMs here i manage them with a TUI here: github.com/AmeanAsad/clau…. I need confidential VMs for a bunch of dev work im doing but it also happens to be a more secure runtime at the same time.

will someone build this / wanna work on this w/ me before @karpathy leaks all the alpha

@karpathy github.com/AmeanAsad/clau…

Expectation: the age of the IDE is over Reality: we’re going to need a bigger IDE (imo). It just looks very different because humans now move upwards and program at a higher level - the basic unit of interest is not one file but one agent. It’s still programming.

@karpathy @nummanali @karpathy rough idea of this here: github.com/AmeanAsad/clau…

@nummanali tmux grids are awesome, but i feel a need to have a proper "agent command center" IDE for teams of them, which I could maximize per monitor. E.g. I want to see/hide toggle them, see if any are idle, pop open related tools (e.g. terminal), stats (usage), etc.

Claude Code teams with tmux is really cool When you run with team mode enabled in tmux, it automatically opens the additional terminal in pane I don't really get my main agent to orchestrate, I chat to them myself CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=true claude