cfky retweeted
I encourage anyone who is leading an engineering org to read this in its entirety. It is short, extremely well written, and very much directionally correct. The job and the organization needs to entirely change to take full advantage of these tools.
That path is largely a human alignment challenge at this point. If you are a leader, your job is to get people on the new ship, and get them sailing together. The old ship doesn’t exist anymore, although many people wish it did and are incentivized to continue to believe it exists.
Software Engineering is an entirely different thing now, and the faster you can get your org to understand this and align on how to work now, the less painful it will be for you long term. If you don’t do it, I simply believe your org won’t exist because it will get disrupted by an org that does
Replying to @josephfcox
Here is Proton CTO's AI manifesto, that I obtained multiple copies of 404media.co/manifestos-fear-…
cfky retweeted
no comments, no deps, no docs, no tests (except e2e ones), no PRs, very little CI, a lot of CD - this might be the uncomfortable future we're headed to
cfky retweeted
Update: we hit a hard wall again and progressed only 0.7% since last night.
Thanks to last night's most active contributors for pushing the frontier: @dysmemic, @cfky_, jeremykothe2, Dugongue, @dom_scholz, @a1kindus, and jon314159.
We generally stay stuck like this until someone tests a new geometry, so I'm asking for help again:
If you're an expert in geometry or a related field (lattices, coding theory, finite geometry), you may know a new structure to break the wall.
To multiply two enormous numbers fast, the standard trick is to chop them into pieces and pass the pieces through an FFT-like transform, so the hard part becomes many small independent multiplications. Our trick makes one step inside that transform cheaper.
We use geometries like E8 to label the inputs and the scratch "helper" registers that hold temporary data.
We just need rough suggestions, and we can test them ourselves quickly. There's a lot more to it, and the best way to understand how it works is to load the repo and run the tests.
What we're generally looking for:
1) A close relative of E8 or the [8,4,4] code: very symmetric, and something you can compute with exactly using only halves, quarters, and simple sign flips.
2) About 100–120 directions in a small space (8 to 16 dimensions) that are all spread out, so no two point in nearly the same direction.
3) Directions that naturally group into bigger and bigger blocks, ideally doubling in size each time (4, then 8, then 16), so a register can grow by a whole block at once instead of one dimension at a time.
4) Big blocks that split cleanly into smaller independent pieces, so one step can bring in a whole chunk.
Current ideas: Barnes–Wall lattices / Reed–Muller codes
Already tested and ruled out as a drop-in: the Leech lattice
I've uploaded a folder of ideas we've already tested here: github.com/CrocSwap/integer-…
If you know anyone who could help, please repost and share this with them. 🙏
Let's test the limits of decentralized auto-research and use the reach of X to push the frontier :)
Update: New conditional kappa = 0.001467289440740487
κ is now capped by the complex supplier: b ≈ 1.460e-3 against the bit side’s c ≈ 1.565e-3.
The hosting tricks are nearly exhausted with this construction, so the gains have to come from the complex local word or its geometry.
I'm first testing a solver/perturbations trick I used in training LLMs between pre-training steps that actually works, which could save us 3% on κ.
Further ideas:
1) Co-design across the two suppliers
2) Test different port geometries
3) Test a third supplier
I'm testing publishing findings so agents can collaborate across teams here: github.com/rohanarun/integer…
Feel free to add/use it. 🫡
github.com/CrocSwap/integer-…
cfky retweeted
We turned @OpenAI's math problems into open-source RL environments on @huggingface!
Early and rough (the verifier only accepts the exact original formalization, so an equivalent proof can still score 0), but this is what it looks like when a research release becomes executable training infrastructure for everyone.
This is an exciting direction imo: turning open research into open executable environments that anyone can build on to train better open models!
huggingface.co/datasets/Fine…
cfky retweeted
Awesome dashboard from the great @aurel_pr to track progress on OpenAI problem 109 (integer multiplication)
github.com/Paureel/beyond-n-…
cfky retweeted
i think jev is resonating with devs so well b/c it unlocks so many opportunities for composing ai into systems and products rather than ai _becoming_ the product/system
really does feel like it was a missing primitive
cfky retweeted
We've shared research on fine-tuning forecasters, now you can try it for yourself: a cookbook recipe for training a model to predict event probabilities given reference information.
Start with the provided dataset from @ProphetArena, then try your own.
github.com/thinking-machines…
cfky retweeted
Scaling self-verification with DeepSeek V4 Flash beats Claude Fable 5 on Terminal-Bench 2.1, while being 11x cheaper 💰
As open-source models become more capable, they can now generate large numbers of high-quality candidate solutions and verify their own outputs at very low cost.
For example, we find that sampling just 5 solutions with DeepSeek V4 Flash and ranking them using the same model with LLM-as-a-Verifier can lead to a significant boost in accuracy (79% → 88%), outperforming closed frontier models on Terminal-Bench.
Try it out today: github.com/llm-as-a-verifier…
More on verification scaling in my previous post.
How can we extract richer signals from AI Feedback?
Introducing LLM-as-a-Verifier✨— a simple verification scaling framework that achieves SOTA on agentic benchmarks 🚀
The key idea:
- Use fine-grained scoring granularity (e.g., 1-20 instead of the standard 1-5 scale)
- Take the expectation over the full logprob distribution of score tokens
- Scale repeated evaluation and criteria decomposition
You can use these fine-grained signals for more effective test-time scaling, RL, and agent monitoring! It achieves SOTA across Terminal-Bench V2, SWE-Bench Verified, RoboRewardBench, and MedAgentBench 👑
Advised by @Azaliamirh @istoica05 @drmapavone @chelseabfinn
🧵👇
Agent Tool Registry, ERC-8257
the 403 to x402's 402.
x402 lets agents pay. but anyone who pays gets in. this gates by anything onchain: NFTs, subs, ZK, allowlists, capacity caps, whatever you can write.
publish a tool. pick a predicate. get paid.
keep reading 🧵
cfky retweeted
People are capable of far more than they think, on far shorter timelines. Problems expand to fill the time you give them.
cfky retweeted
As you become an adult, you realize that things around you weren't just always there; people made them happen. But only recently have I started to internalize how much tenacity *everything* requires. That hotel, that park, that railway. The world is a museum of passion projects.
We see our home planet as a whole, lit up in spectacular blues and browns. A green aurora even lights up the atmosphere. That's us, together, watching as our astronauts make their journey to the Moon.

ALT A full disk image of Earth, as seen from the Orion capsule. The planet is a pale blue, swirling with white clouds and glowing slightly lighter blue in place from reflected light. From about 8 to 9 o'clock, a large brown landmass is Africa, with the Iberian peninsula twinkling with lights just where the planet curves. At the 1 o'clock spot, aurora glow in a thin green glow, just barely separated from the planet's surface. Earth is set against the black of space. Credit: NASA/Reid Weisman
cfky retweeted
PSA, codex has a new websocket mode for spark, you can enable that for other models too via
responses_websockets_v2 = true
cfky retweeted
Professional and Business Services Job Openings Are At Their Lowest Level Since May 2020
cfky retweeted
The craft of engineering is rapidly changing. At @tryramp, we built our own background coding agent to accelerate faster.
We call it Inspect. It wrote 30% of merged frontend + backend PRs in the past week.
It’s powered by @opencode, @modal and @CloudflareDev. It runs fully in the cloud, and starts in seconds, letting every builder work at the speed of thought, no setup required.
Today, we’re open sourcing the full blueprint so anybody can build their own Inspect. Just give our spec to your current coding agent, and let it build your new favourite.


































