Jeff Smith

Jeff Smith

@jeffsmith.tech

Building C Proof: https://cproof.ai/ Researching and writing: https://www.jeffsmith.tech/

726 followers2.7K following332 posts

View full thread
Jeff Smith
@jeffsmith.tech

Last third is my own engineering work. SHARe-KAN compresses pre-trained KAN inference for multi-expert deployment. Brill Agents compiles a trained KAN out of existence at inference and leaves native BEAM pattern-match clauses behind at sub-millisecond latency.

Jeff Smith
@jeffsmith.tech

Both are bets on the engineering substrate the structuralist research program needs. The compressions of intelligence won't be written without it. 55 minutes. Block out the time.

View full thread
Jeff Smith
@jeffsmith.tech

Robert Ronan does the math. Hilbert's 13th problem, the Kolmogorov-Arnold representation theorem, B-splines, the architectural reframing that puts the learnable nonlinearity on the edge. He carries the precision the rest rests on.

Jeff Smith
@jeffsmith.tech

Today the field mostly puts structure outside the model, through prompts and agent frameworks. KANs put structure inside the model. The bet is on the second path. Most of the field hasn't engaged with it. This episode is my attempt to fix that.

View full thread
Jeff Smith
@jeffsmith.tech

There's also some personal history in this one connecting back to Papers with Code and FAIR, but mostly this is about what's happening now and why open infrastructure for RL matters more than ever.

Jeff Smith
@jeffsmith.tech

To do remarkable work you must be working in a manner where your work can be remarked upon. Stealing that from Cal Newport but it fits perfectly here. Train in the open. Let people see the results.

View full thread
Jeff Smith
@jeffsmith.tech

Benchmarks are saturating. The hard stuff is long horizon, non-stationary, requires real adaptation over time. OpenReward has 330+ environments and a fully open standard. The community building complex worlds to test these capabilities is growing fast.

Jeff Smith
@jeffsmith.tech

The hackathon had formal verification environments next to elderly care simulations next to game AI. All on the same platform. The diversity of ideas coming out of London's AI community right now is something else.

View full thread
Jeff Smith
@jeffsmith.tech

When they turned on chat between agents in Catan, things got weird. Some models started fabricating sob stories to negotiate better trades. The other agents just accepted it. Emergent deception nobody programmed.

Jeff Smith
@jeffsmith.tech

The metaphor running through the episode is the Olympics. What if athletes trained in the open every day and could watch each other's methods? That's RL right now. The arenas are open. The records are falling. And you don't need to be at a frontier lab to compete.

Jeff Smith
@jeffsmith.tech

Julie Huang and Khalid Altalib won the Complex Worlds Hackathon. Built a pipeline that converts any Python game into an RL environment on OpenReward. Catan, Overcooked, Diplomacy, and their own business sim. Six hours. #RL #OSS

View full thread
View full thread
Jeff Smith
@jeffsmith.tech

And if you have other ideas about what to do about expertise now that the 🤖s are writing all the code, I'd love to hear how you're tackling it.

View full thread
Jeff Smith
@jeffsmith.tech

Still, there's such a thing as consequential computing. Code that touches money, health, human life. Someone still needs to have climbed the mountain and developed the expertise.

Jeff Smith
@jeffsmith.tech

The SWE role is getting harder *and* more interesting. The days of body shops are over. This is rapidly becoming a job for people who can legit bring a huge amount of leverage, because we just don't need to push a bunch of boiler plate around with human hands.

View full thread
Jeff Smith
@jeffsmith.tech

The @opensource.block.xyz PR is maybe my favorite. @Jack Dorsey is the author. The agent picked "approximately 4 characters per token" as a cost heuristic and capped model memory at 20. Gets into authority in code review. #vibecoding

Jeff Smith
@jeffsmith.tech

Lots of knock-on consequences: PR review as cargo culting. Clean-room reimplementation to license wash and take over control. People are ready to wildly change the rules of the SDLC within and beyond #OSS.

View full thread
Jeff Smith
@jeffsmith.tech

PEAK asks totally different things depending on the repo. The #scipy questions about FITPACK and rank deficiency are a world apart from the #openclaw questions about session state architecture.

Jeff Smith
@jeffsmith.tech

The core research (first shown at @qconferences.com London): every repo has rules nobody documented. You only know them because you've been there for years. We pull them out computationally from commit history. PEAK turns them into questions. #opensource #airesearch

View full thread
Jeff Smith
@jeffsmith.tech

The video also covers xz Utils, the LiteLLM supply chain attack, the openclaw PR flood. But the punchline is optimistic. Projects are writing contribution guidelines that talk directly to agents and it's working. The clankers are steerable. The slop spike is self-correcting.

Jeff Smith
@jeffsmith.tech

This is the first episode of a new series called Iterative Refinement from my company Burnin. Each episode is a complete deep dive into one thing we built. How it works, what failed, what we learned. Long form, minimal editing, just a camera and the work. www.youtube.com/channel/UC4D...

youtube.comIterative RefinementIterative Refinement is a video series from Burnin (burnin.ai). Each episode is a complete deep dive into one thing we built. How it works, what failed, what we learned. We cover AI evaluation, agent ...
View full thread
Jeff Smith
@jeffsmith.tech

Tried to detect suspended accounts. Tested merge rate, network centrality, TF-IDF, LLM scoring across 31K Gemini calls. Best result: AUC 0.619. Nearly chance. If someone already passed human code review their profile looks like everyone else's.

Jeff Smith
@jeffsmith.tech

Stripped the graph out. Accuracy went up. Account age added nothing. Every attempt to make the model smarter made it worse or flatlined. v3 is one line. merged / (merged + closed). Sometimes the simplest thing wins.

Jeff Smith
@jeffsmith.tech

Started with a complex graph model. PageRank over a bipartite contributor network, language multipliers, anti-gaming penalties. Scored famous contributors beautifully. Scored the people who actually need scoring terribly.

Jeff Smith
@jeffsmith.tech

This is actually a massive result that people are really all sleeping on. We can just start over from zero and, with the right language design, build a programming language that the models can write perfectly without a training corpus.

Alasdair Allan@alasdairallan.com

Expanded the Vera benchmark to 6 models, 3 providers. Kimi K2.5 got every Vera problem right while only managing 86% in Python and 91% in TypeScript. Flagship average Vera and Python level at 93%. Yet there is no Vera in their training data, no examples on GitHub or Stack Overflow. veralang.dev #AI

Jeff Smith reposted

Michael Lobel
@mlobelart.bsky.social

I have nothing useful to add on the general state of things right now, but I can offer a thread of canine portraits by 18th-century French painter Jean-Jacques Bachelier, who was also director of the famed Sèvres porcelain factory. First up: Mimi, the favorite dog of Madame de Pompadour, 1762

Jeff Smith
@jeffsmith.tech

In the space of an afternoon, you can see thousands of devs around the world benchmarking their current vendor's competitor, mid outage. Just one more sign of the insane pace of change with agentic AI.

Jeff Smith
@jeffsmith.tech

Back in the day, being a dev used to be like being Ludwig from that David Mitchell show. You should just pace and whiteboard and ponder until you solved the puzzle. And you felt really smart when you did. It was a big part of the appeal for a lot of us.

View full thread
Jeff Smith
@jeffsmith.tech

Back in the day, being a dev used to be like being Ludwig from that David Mitchell show. You should just pace and whiteboard and ponder until you solved the puzzle. And you felt really smart when you did. It was a big part of the appeal for a lot of us.

Jeff Smith
@jeffsmith.tech

But with agentic coding, basically being a dev is just Starcraft with shittier graphics. You barely even notice all of the wins, because the model is solving all of the puzzles. You just wind up typing as fast as you can. It's pure video games.

Jeff Smith
@jeffsmith.tech

So, you don't feel like Ludwig when you're done. You just feel like you played as fast as you could but the game can still be played so much faster and there's probably a teenager in Korea who could mop the floor with you.

Jeff Smith reposted

Andri Óskarsson
@andri.dk

The new “This could have been an email” is when someone does interesting research and the only way to consume it, is on YouTube “This could have been a blog post!”

Jeff Smith
@jeffsmith.tech

Which is why, although all John Luther Adams music will put you to sleep, all John Adams music will disturb your rest forever.