←
Cong Lu851 posts
Cong Lu avatar

Cong Lu

@cong_ml

Automating Scientific Discovery @Recursive_SI. Prev: @GoogleDeepMind | @UBC_CS | @UniofOxford | @SakanaAILabs | @Waymo. The AI Scientist | Genie | SIMA | Gemini

◷ Joined October 2019▧ 77 Photos and videos
Cong Lu avatar
Cong Lu@cong_ml·Oct 27

1/ You can now run a research swarm entirely on your Mac, model included. Eight agents share one local 40B Flint model on a 48 GB M4 Pro: writing code, testing ideas and building on whatever works. Give them a problem and a score to beat. Here's how 🧵

0:16
Cong Lu avatar
Cong Lu@cong_ml·Nov 11

It was a lot of fun pointing our Eureka Machine at alien optimisers 👽👽👽, glad we could demonstrate some of the insanely wild creativity of AI there Even more fun contributing together with some old friends at @SakanaAILabs @RobertTLange ;)

Optimization research has always sought simple, elegant learning algorithms that are homogeneous across all architectures. Over the last two months, we've seen a rise of optimizers that are increasingly capable, but also increasingly complex, and ad hoc. We call these alien optimizers: learning rules found by search and metalearning instead of designed by hand. We think this trend will accelerate, producing optimizers that co-evolve with architectures and are far more sample-efficient, at the cost of complexity. w/ @akshayvegesna @cs_serdar

Post attachment
Robert Lange avatar
⟳ Cong Lu retweeted
Robert Lange@RobertTLange·Nov 3

What happened to aider? I still remember when @cong_ml first showed us the OG precursor to Claude Code/Codex/Pi/OpenCode. This was back in mid-2024, when the dev standard was still simple autocomplete in VS Code. Aider was really ahead of its time. It already had a TUI interface, a simple-to-use Python SDK (which we leveraged in the AI Scientist v1), voice mode, and worked really well given the state of LLM tool-calling capabilities back then. The docs still contain some really clever engineering details for context management and codebase mapping. Ah and Polyglot, a coding benchmark based on exercism tasks, helped kickstart a lot of coding agent self-improvement projects. Hitting the right trade-off between costs and improvement signal quality. In November 2026, the repo has ~50k stars and no commits on main within the last 6 months. Really hoping history will remember this crucial stepping stone and @paulgauthier's immense efforts. Open source is pretty amazing 🥰 Code: github.com/Aider-AI/aider

Post attachmentPost attachment
Demis Hassabis avatar
⟳ Cong Lu retweeted
Demis Hassabis@demishassabis·Oct 27

Very excited to share Genie 4, our incredible interactive world model, bringing together real-time physics, reasoning, and audio! Simulations within simulations, makes you wonder just how deep the rabbit hole goes... I have a feeling we will be finding out soon enough...

This is Genie 4, our most realistic and interactive world model yet. 🌐 Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.

2:27
Josh Tobin avatar
⟳ Cong Lu retweeted
Josh Tobin@josh_tobin_·Oct 27

I continue to be surprised and impressed at how much autoresearch you can do on your laptop these days thanks to Flint

1/ You can now run a research swarm entirely on your Mac, model included. Eight agents share one local 40B Flint model on a 48 GB M4 Pro: writing code, testing ideas and building on whatever works. Give them a problem and a score to beat. Here's how 🧵

0:16
Tim Shi avatar
⟳ Cong Lu retweeted
Tim Shi@timshi_ai·Oct 27

swarms of Flint on your mac

1/ You can now run a research swarm entirely on your Mac, model included. Eight agents share one local 40B Flint model on a 48 GB M4 Pro: writing code, testing ideas and building on whatever works. Give them a problem and a score to beat. Here's how 🧵

0:16
Tim Rocktäschel avatar
⟳ Cong Lu retweeted
Tim Rocktäschel@_rockt·Oct 27

Great news for doing science on autoresearch swarms at home.

1/ You can now run a research swarm entirely on your Mac, model included. Eight agents share one local 40B Flint model on a 48 GB M4 Pro: writing code, testing ideas and building on whatever works. Give them a problem and a score to beat. Here's how 🧵

0:16
Cong Lu avatar
Cong Lu@cong_ml·Oct 27

6/ A year from now, what will a research swarm on a laptop be able to do? My earlier local autoresearch demo: x.com/cong_ml/status/2112966… Code and examples: github.com/conglu1997/local-…

1/ From OpenAI's Navier–Stokes work to AlphaEvolve's faster algorithms, AI research is getting very real. Now, with Recursive Flint, we can experience that magic locally on our own laptops 💻 The whole research loop runs locally with a selection of tasks! Code + demos 🧵

Cong Lu avatar
Cong Lu@cong_ml·Oct 27

5/ To see how much faster the swarm could work, we ran one and eight agents for six hours each, twice per task. Across three tasks, eight agents averaged 2.3× as many valid experiments per research hour. Here's how the solutions improved:

Post attachment
Cong Lu avatar
Cong Lu@cong_ml·Oct 27

4/ So each agent also gets a different instruction: refine the method, try another algorithm, simplify it, or question an assumption. Agents rotate through these on every attempt. It's a simple version of the prompt variety in systems like AlphaEvolve.

Post attachment
Cong Lu avatar
Cong Lu@cong_ml·Oct 27

3/ The agents collaborate through one shared result: the best program so far. When one agent beats it, the others start from the new best on their next attempt. But if everyone starts from the same program, they tend to try the same things.

Cong Lu avatar
Cong Lu@cong_ml·Oct 27

2/ All eight agents share one copy of Flint in llama.cpp, which batches their requests together. On an M4 Pro, eight requests at once produce 1.9× the tokens per second of one at a time. Each agent writes more slowly, but together they write almost twice as much.

Post attachment
Cong Lu avatar
Cong Lu@cong_ml·Oct 22

5/ Autoresearch helped us develop Flint. Seeing Flint contribute useful research itself is an encouraging early step. We're starting to explore how far that loop can go, and how to develop it safely. Meet Flint: x.com/cong_ml/status/2112949… Try it locally:

1/ From OpenAI's Navier–Stokes work to AlphaEvolve's faster algorithms, AI research is getting very real. Now, with Recursive Flint, we can experience that magic locally on our own laptops 💻 The whole research loop runs locally with a selection of tasks! Code + demos 🧵

Cong Lu avatar
Cong Lu@cong_ml·Oct 22

4/ Flint finished close to historical Haiku 4.5, with Sonnet 4.6 still ahead at the same experiment budget. Our 40B model uses just 2.5B parameters per token. We didn't expect it to get this far on a research task we'd associated with powerful proprietary models.

Cong Lu avatar
Cong Lu@cong_ml·Oct 22

3/ Here, Flint mainly tuned hyperparameters for a very specific training budget. It explored batch size, depth, learning rates and schedules—testing combinations and building on what worked. The progress came from adapting the recipe to the constraints.

Post attachment