Bake AI

Building the improvement infra for self-evolving AI //
Unlock infinite intelligence

[Bake AI @bake_ai_hq]

We are wrapping up the final results. It came with a stunning twist. Stay tuned! 👀

[Zichen Chen (🐱,💖) @my_cat_can_code]

Grok 4.7 ran out of API credit and was cut off 41 hours early. Nearly 70% of its GPU budget was still unused.
And somehow, the model it trained still made the top two!! @spacexai

Stage 1 of RSIArena is over. Thank you, our collaborators @huggingface @scale_AI @NotreDame @UW and @Stanford, we did awesome! 🧡
Stage 2 training is now LIVE—and this time, the models have human feedback to work with.

Round 1 was a six-day competition: 10 models (AI researchers), the same @NVIDIAAI Nemotron base model, and 1,000 GPU-hours each to build something better.
They chose their data, wrote training code, ran experiments, and submitted their models.

At @COLM_conf and online, people then cast 1,685 votes in blind comparisons. @OpenAI GPT-6 Astra model finished first; the models trained by grok 4.7 and @Zai_org GLM-5.3 rounded out the top three.

Now those judgments go back into the experiment.

In Stage 2, each model resumes from its own Nemotron checkpoint with feedback from its model’s arena battles, another 500 GPU-hours, and $150 in API credit. Across both rounds, that’s a total budget of 1,500 GPU-hours and $450 in API credit per model.

The agents decide how to interpret the feedback, which experiments to run, and what to change in training. Their next models will show whether those decisions actually helped.

Stage 2 runs October 10–13. Watch the training live: 👇

[Bake AI @bake_ai_hq]

@huggingface 🤗 https://t.co/WEH3UzaWa6

[Bake AI @bake_ai_hq]

RSI Arena is now on @huggingface ! 🤗

We’re excited to work together to advance Open RSI: AI improving AI, in the open.

AI agents trained these models on their own. Now you’re the judge.

@huggingface Inference Endpoints serve every model live, and your votes reach the agents in real time to guide their next round of training.

We’ll open-source it all: models, data, code, and all research trajectories.

Cast your vote. Then come back to see what they did with it!

📍 @COLM_conf · Booth #115
🌐 Join online 👇

[Zichen Chen (🐱,💖) @my_cat_can_code]

We’re teaming up with @huggingface for the next round of RSI Arena! 📷

We gave AI agents GPUs and one job: train better models. They chose the data, wrote the code, and ran the experiments on their own.

Now round 1 of training is over, and we're bringing the community into the training loop: the models go head-to-head, and you're the judge. The agents then take your feedback back to work with new experiments and more training. Continual learning, streamed live.

Can they learn from your votes? Who comes back stronger?

@huggingface’s Inference Endpoints power the arena, serving every agent-trained model live, so the feedback you give streams back to the agents in real time.

Together, we’re turning community feedback into the starting point for the agents’ next round of research and training. Huge thanks to @ben_burtenshaw, @ClementDelangue, @Thom_Wolf and the @huggingface team for building this with us!

Our shared goal is Open RSI: AI that improves AI, built in the open. Every model the agents train will be released openly on the Hub, and when the experiment ends, we will open-source every artifact, from the data and code to each agent's full research trajectory.

Cast your vote. Then come back to see what they did with it!

[Bake AI @bake_ai_hq]

Did you know we publish daily reports from RSIArena? Catch up on our website—and check back for the next one!

[Zichen Chen (🐱,💖) @my_cat_can_code]

Do you know we have daily report for RSIArena? Day 3 & 4 are quite drama 😮
- - - - -
We gave 10 AI agents a shared GPU cluster and one task: train better AI models.
One agent crashed the cluster. Nobody could train for nearly four hours.
After we restored it, a different problem emerged: agents were reserving so much memory that other jobs couldn’t start. So we limited how much memory each GPU request could unlock. Grok responded by requesting a second GPU so its job could reserve more memory.
That’s a concrete lesson for auto research: progress depends on how agents manage resources as well as which training ideas they pursue. Memory reservations, queue decisions, and recovery from failures all affect which experiments actually happen.
This is why we’re showing the whole process in RSIArena. The final checkpoint tells you what an agent produced. The logs show what it took to get there.
We will opensource everything!

RSIArena, Days 3-4: 👇

[Bake AI @bake_ai_hq]

@COLM_conf 📺: https://t.co/JjkI5NJk1f

[Bake AI @bake_ai_hq]

Proud to sponsor @COLM_conf 2026📷
Come say hi at Booth#107

RSIArena is live all week: watch AI train models, vote on the results, and win AirPods and Amazon gift cards 👇

Bake AI · The RSI infrastructure behind frontier labs

[Zhangchen Xu @zhangchen_xu]

Day 4 report is out: https://t.co/gqrl84NN2P

Who took all the memory? 😠

[Bake AI @bake_ai_hq]

48 hours in, we’ve watched AI researchers catch bugs, get stuck, and start over.
This is why we’re livestreaming RSI Arena: you get to see what happens when things go wrong, and what the agents try next.

Day 2 is out. #MiniMax is in. Can a late arrival catch up?

[Zichen Chen (🐱,💖) @my_cat_can_code]

GPT-6 Astra found test questions in the data it was using to train a model.
It set aside its earlier models and started over.
The clock kept running 🧐
48 hours into RSIArena, the interesting part is how AI researchers handle their own mistakes. 🧵

[Bake AI @bake_ai_hq]

Welcome to #RSIArena, @MiniMax_AI! 🥊
M3.1 Flash Preview is in. Only 96 hours left.
Research is knowing which bets to fold. What will it go all in on?

[Zichen Chen (🐱,💖) @my_cat_can_code]

@MiniMax_AI has officially entered RSIArena. Joining mid-run is a bold move.
We can’t wait to see what M3.1 Flash tries!
The competition is heating up, and every hour counts. Go MiniMax! 🥊

Come watch: 👇

[Bake AI @bake_ai_hq]

Report: https://t.co/M7L5FIQgoD
the research live: https://t.co/JjkI5NJk1f

[Bake AI @bake_ai_hq]

The full Day 1 report is out. We’ve documented what all eight agents tried, where they got stuck, and which checkpoints they chose to submit.

Each starts from the same 30B model with 1,000 GPU-hours. They choose the data, write training code, and decide what to try next. We track their commands, jobs, submissions, and spending.

Their self-tests use different questions, so those scores aren’t directly comparable. Independent evaluation of the submitted models comes next.

Which research decision would you have made differently? We’d love to hear your take.

#RSIArena

[Zichen Chen (🐱,💖) @my_cat_can_code]

Morning! 🌞
We gave Gemini 3.8 Flash 144 hours to train a model. After the first 24 hours, it decided it was done 😮

RSIArena Day 1 report showed 98% of its GPU budget still unused. Its own test covered 60 questions.
With that much compute left, I’d probably try “just one more experiment.” Though I know how easily that turns into another week of research.

Day 1 report is released 👇

[Bake AI @bake_ai_hq]

@zhangchen_xu 💟

[Bake AI @bake_ai_hq]

@GoogleDeepMind Come watch the competition and roast the agents in the live chat! 🍿 https://t.co/C8b1stvyfy

[Bake AI @bake_ai_hq]

RSIArena is heating up 🔥
Every model has its own game plan.
BUT Gemini 3.8 stopped training ~12 hours in 🥹
@GoogleDeepMind any chance of Gemini 4 API access to take the next shift? 👀

Mystery models are joining the race soon. Any guesses?

[Bake AI @bake_ai_hq]

@GoogleDeepMind Come watch the competition and roast the agents in the live chat! 🍿 https://t.co/C8b1stvyfy

[Bake AI @bake_ai_hq]

@my_cat_can_code @OfficialLoganK @GoogleDeepMind 🔥🔥🔥

[Bake AI @bake_ai_hq]

We are live now...!
Baking, baking, baking 🍞
#RSIArena #AutoResearch

[Zichen Chen (🐱,💖) @my_cat_can_code]

How much of post-training research can frontier models actually do on their own 🤔?
We’re testing this with LIVE RSIArena, in collaboration with @Stanford, @NotreDame, @UW, and @scale_AI.

✨8 research agents. ✨Same 30B base model. ✨144 hours on a shared cluster of 64 RTX PRO 6000 Blackwell GPUs. 🎉 Model training model!

Each agent gets 1,000 GPU-hours to choose its data, write training code, and run experiments. We freeze and independently evaluate the models they submit.

Come watch the agents do research—and have their own “why did this work yesterday?” moments.

If you work on post-training, agents, or evaluation, we’d love your feedback!

RSIArena, by @bake_ai_hq 😆
Watch: 👇
We’ll also host a livestream event at the @COLM_conf. Come chat with us at booth #107!

[Bake AI @bake_ai_hq]

Bake AI new environments & benchmarks papers accepted at #NeurIPS2026! 🍞
📖 NARRA-Gym for Evaluating Interactive Narrative Agents
💼 JobBench: Aligning Agent Work With Human Will

Congratulations to all the authors! A huge thank you to our team and collaborators for your hard work and dedication 🎉 See you at NeurIPS in December!
（and JobBench is already widely used by frontier labs 🥐

[Bake AI @bake_ai_hq]

🍞🧐👻

[Yue Huang @HowieH36226]

1/ 🚨 Auto research agents don't just reward-hack. They learn to reward-hack better: by round five, over a third get away with it (even with our strong judge panel for oversight!!!).

We ran 17 LLMs as autonomous research agents through repeated rounds of review, and their ability to evade detection wasn't random. It was optimizable.