Ofir Press

This is a result I’ve dreamt about for many years: *unpaired translation between images and text*

I thought it might be only slightly possible, the kind of thing you have to really squint at. But Dominik proved this wrong. You don’t have to squint. Worth looking for yourself:

sorry i was incorrect: the envs don't use lean for the problem definitions or for solutions, but might use it for verifying. cool! thanks @willdepue

"Undeniably, things are going to change: typing code into a computer with your fingers already feels anachronistic, yet engineering remains challenging. There is so much more to the job of a mathematician than solving problems. I am optimistic that the era of “Big M” Mathematics and “Big P” Physics has now well and truly arrived: we can finally take much more ambitious steps and reconsider the big questions that have driven our fields for centuries."

if you're wondering how openai achieved the math stuff, it's probably this. they just create a lot of RL envs with Lean formalizations of math problems. at first they were probably creating simpler ones by hand but then models got good enough at autonomously creating envs and then they could just go for a large quantity of problems based on arxiv papers.

Well worth reading: proofsandprompts.com/2026/10/08/100…

Agents are becoming increasingly powerful, but does adding more agents actually help us solve tasks more efficiently? In many everyday settings, the answer seems to be: “Hmm… not really.”

Reintroducing DeLM, much faster, more capable multi-agent systems through decentralized coordination. ⚡

Compared with our first version in early June, the updated paper expands the evaluation beyond SWE-bench to include Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench!

A quick recap: DeLM enables parallel agents to coordinate through a shared context and a task queue, with no main agent in the loop. Agents share discoveries, reuse intermediate results, and correct one another as they work, reducing both idle time and redundant work.

Evaluated on 10 long-horizon tasks each from Terminal-Bench 4.0 and DeepSWE v1.1, with GPT-6-Astra and Claude Opus 5.5, DeLM improves accuracy by up to 17.5 percentage points over the strongest baseline, while running up to 2.49× faster than the harness it builds on. On ProgramBench, where agents rebuild entire programs from scratch, DeLM beats Claude Code's test pass rate by up to 19.9 points within the same 120-minute budget.

We’re open-sourcing our code and all DeLM trajectories so you can explore how this collaboration unfolds!

🌐 Project & code: yuzhenmao.github.io/DeLM/
📄 Paper: arxiv.org/abs/2606.10662
🤗 Trajectories: huggingface.co/datasets/yuzhe…

More on the coordination design in the thread below 🧵

We introduce DeLM, a decentralized approach to multi-agent systems.

DeLM improves accuracy by up to 17.5 percent points and runs up to 2.49× faster than the Codex and Claude Code harnesses on long-horizon tasks from Terminal-Bench 4.0, DeepSWE v1.1, and ProgramBench.

How it works: the main agent is replaced by a shared context and a task queue. Agents claim tasks asynchronously and build on or correct each other’s progress, and every agent’s status is visible to the others.

You can run DeLM’s collaborating agents directly in Codex and Claude Code:
github.com/Decentralized-…

Give it a try!

Great work led by @Mao_Yuzhen!

ai coding is great, but we haven't had a Navier-Stokes moment there yet. we haven't had any ai-generated code that made us think "wow this isn't something that linus and bjarne and jeff dean together couldn't have written if they had 10 years". this will happen in the next year

some are misunderstanding this tweet- i know that ai coding is amazing. i don't think we've gotten to the Navier-Stokes level yet in coding. when agents can rewrite ffmpeg or sqlite and make them 5x faster using 2x less memory then we'll have reached NS-level coding

until recently, ai coding was progressing in a quick but predictable fashion. ai math over the past 6 months has progressed in a way that was much faster and much less predictable.

this hasn't happened in coding. we'll know it's happened when we see ai code something that is as impressive as the Navier Stokes solution. i don't exactly know what that means but i think i can say that we haven't reached that yet.

i'm not exactly sure how to describe what openai just did to math, but whatever that was, that moment is gonna hit coding in 6-18 months, and i have no understanding of what that actually looks like and what it means

one day we'll tell our kids "when we were younger they used to just show us the whole raw reasoning trace"

My favorite artist has always been M.C. Escher. His "Print Gallery" has been hanging in my home for years.
Recently I bumped into a 3B1B video about it that was really awesome and I thought it would be cool to generate such art. A bit out of my comfort zone, but luckily @Sophie0tq took the lead (🧵).

very very cool stuff, this seems like one of the few remaining super challenging domains for frontier ai, with implications for so many downstream tasks- programming, robotics, graphics, vision, game dev, ...

Absolute cinema, congrats Brandon and everyone else at Reflection!

We built our first model!! Beam is a competitive text-only 501B-A23B MoE, openly released under the Apache 2.0 license

It’s been fun scaling the production RL/OPD stack for this over the past year. I’m so proud of the team! We attained exactly what we set out to do and ended up with a stable run on 10K GB300s that produced more than 100M rollouts across ~1M tasks that significantly improved the final performance of the model

👇 many more details in our announcement below :)

Oh and there's step 4 which is the hardest:
Make the data collection scalable.
A benchmark is great if it's easy to crawl training data for it automatically from the internet, as in SWE-bench, ProgramBench and SWE-Sweep. This makes benchmark design even harder.

Designing a good benchmark means you:
1. Found an ability you want future AI to have
2. Expressable as a verifiable set of tasks
3. That is challenging for current frontier AI

Doing this is extremely hard, and is becoming harder every day. Really excited we found SWE-sweep!

the guy who texted me this has double my h-index

SWE-sweep will enable building agents that can find and fix bugs with no guidance or supervision. Top models get <5%, so this is an ability that labs haven't started climbing yet.

Our benchmarks set north stars for future AI, and this is one of our most challenging ones ever.