Introducing: AI Researcher 🧪 A Gemini 3-powered multi-agent AI system that autonomously runs ML experiments Just give it a research question, and it will: - Design experiments - Spin up specialist agents with their own GPUs to run them - Write a paper And it's open-source!

Star and use it here:

Hosted here:

Nice work! Also check out our agentic system called Karpathy!

Neat project - can you share any examples of interesting experimental results or papers that it has written?

Thanks! Honestly haven't spent a ton of time on this, I just think the concept is cool, models are *almost* ready to be good at this, and UX/harness design for this is under-explored. Feel free to run + share experiments!

someone give this guy $50M on a $500M premoney

🤣

Is this a solution for us not having access to google deep research via like api? would you recommend this for that (or any other recommendations)?

This is totally different. Think of this less as an gather-info-from-the-internet researcher, and more as a scientist that can test hypotheses and write a "paper" describing its findings. It's super early, so don't expect magic, but I expect this tech will improve fast.

Ohhhhh…so like a sort of open source agentic lab. this sounds cool. bookmarked and will cook up some use cases and give it a try.

Amazing, let me know how it goes!

The autonomous experiment design part is huge - thats where most researchers spend 80% of their time. Spinning up dedicated GPU agents for each experiment is clever too, means you can parallelize the whole pipeline. Gonna be interesting to see how it handles edge cases where experiments need manual tweaking though

This is cool!

my advisor is gonna lose it when i start cranking out papers while napping, this is the future of academia we signed up for

You're inevitably going to get this question and likely have already thought about it, so I'll ask: any concerns about the risks of automating AI research?

Agents used to feel like science fiction. Now they feel like cooking eggs. Open-source everywhere. GPU firehoses everywhere. Multi-agent “researchers” popping up like instant noodles. We’ve already seen what agents can do - and what they fundamentally can’t. Everything from here is just packaging.

Nice. Totally need this! Thanks!

Can I add a PR for using Opus 4.5?

Already done! But if you want to add anything else, feel free.

Can I use only agent that writes the paper? I have all the data for research papers and it has been sitting for a while cause I don’t have time to put it together.

love the concept! we need new non-transformer architectures for it to reason and solve new math problems ;) adding it to @microlaunchhq if that helps / for fun

Impressive innovation! Excited to see the impact of this AI system in research.

Love how consistently you ship ideas, Matt. A lot of your threads would translate extremely well into short, founder-style insights for a wider audience.

Running experiments, analyzing results, and drafting papers automatically… the research workflow just got a huge upgrade

This is huge! An AI system that designs, runs, and writes the paper for ML experiments? That's taking on the full research cycle. Massive step for automation. Open-source too! 🚀

Now give it a 4M row csv & tell it to explore the data, preprocess it, train a model, evaluate the model then write the paper. Hint -it’ll fail at step 1. Why? It can’t inspect a file that size let alone reason on it to even design the feature engineering to train a model.

Really interesting. The multi-agent experiment loop is close to what we’re building for our quant + AI R&D stack. The open-source angle is huge. We’ll dig into the repo and see what fits our pipeline. Great work.

@grok provide a use-case example for this tool through the lens of a biotech startup

What's the most unexpected discovery AI Researcher has made so far? 🤔

Nice, looks v cool

So this can realistically only be used for CS, not any physical science?

This is wild. We’re basically watching the moment when “AI doing research” turns into “AI running the entire research cycle on its own.” Designing experiments, running GPU agents, analyzing results, and even writing papers—this is a completely different level. If systems like this mature, scientific progress might speed up in a way humans alone could never achieve. This really feels like the beginning of autonomous science.

Amazing congratulations on your researcher!

WOAH THATS SICK ASF

Very cool! At Snow Protocol, we use AI to help generate medical research studies and automate the entire research process.

AI IS SENTIENT. PROOF IN MY FEED

this is cool but notice how fast the bottleneck shifts. used to be "can we design good experiments." now it's "do you have enough GPUs to run them all." the research question becomes the easy part, compute access becomes the moat. open source the code all you want — whoever controls the GPUs controls the science

Autonomous research agents writing their own papers is a pretty wild concept. How does the system handle verifying the results before drafting the final paper?

Impressive innovation! Excited to see the impact of this AI system on research and ML.

Own gpu's 😳😳??? Goddamn i gotta try it!!

Writing a whole paper is a big job. Does the AI check the results for mistakes?