Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. github.com/karpathy/nanochat… All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

Ilya Sutskever: We are no longer in the age of scaling, we are back to the age of research

We let the Internet down today. Here’s our technical post mortem on what happened. On behalf of the entire @Cloudflare team, I’m sorry. blog.cloudflare.com/18-novem…

I vibecoded this neural network visualization for my students and open sourced it. It shows a simple MLP trained on MNIST handwritten digits at several training steps. The visualization is using @threejs and it comes with training code in @PyTorch . Link repo 👇

EdgeTAM, real-time segment tracker by Meta is now in @huggingface transformers with Apache-2.0 license 🔥 > 22x faster than SAM2, processes 16 FPS on iPhone 15 Pro Max with no quantization > supports single/multiple/refined point prompting, bounding box prompts

Attempting the Impossible: Teaching a Neural Network to Predict Pi's digits Here's what we're doing, in the spirit of futility mixed with curiosity: We grabbed 10,000 digits of Pi and gave each one a unique binary "fingerprint, like assigning a bar code to every position (N choose K encoding) like 11111111100010000000,5 (4th digit wihch is 5) Then we built a ridiculously overengineered neural network (1024-neuron hidden layers, 7 residual blocks, fancy learning-rate schedules, the works) and asked it: "Here's the bar code for position 10,857. What digit comes next?" (predicting the next 1000 unseen digits) it pretty much memorizes the training data but guesses at chance (10%) for unseen digits :( although Pi's digits are mathematically proven to be randomly distributed, one can still try right? But I had to see it with my own eyes 😝 I will report back if a miracle happens