Gary’s next surprise will be self-updating AI which will find the next breakthroughs. After GPT-5 and Llama-4, no one is shrugging in dismay, turning off their gpus, and going home.

Multiple frontier labs indicated that they have witnessed AI self-update. I have witnessed AI self-update from an open source repo. Don’t miss the exponential because of a couple of underwhelming launches. Product and research are often working on different timelines.

What do you want to know?.

Having way too much fun pretraining with this repo: While also creating ideal synthetic pretraining data from this paper:.

As a CEO of an AI company it’s a weird thing to have your most valuable people sitting around nursing training runs. We often optimize the time reading papers and playing with new tools, but it’s super weird to have multiple PhD devs just shooting the breeze so often.

Can we have o3 back?.

Today

Just read a proposal on optimizing AI to be maximally persuasive and it’s actually not that hard. This is going to get weird.

I can’t explain why Claude code is better than Cursor, but it just is.

“ Improving itself”. Q4 2025 won’t be about agents, but about autonomous self-improvement.

Fast takeoff 2026.

Pulitzer Prize for this man.

with Qwen3-0.6B as the base model. The previous model we used was 10x larger. This one can fit on edge devices. Welcome to a new era. It's still running.

doing research as to why it improved the way it did, suggesting how to improve itself the next time, and then modifying its own code based on this research. We have now achieved a new best score on Swahili>Rwila machine translation (an extremely low-resource language).

I chose Qwen3-0.6B as the model which is small enough to run locally on most devices, and one of our datasets of an extremely low-resource language in Tanzania. The results?. The AI is autonomously improving its own architecture by creating an experiment, measuring its results,.

This paper got some negative comments because there’s some hype language in it . This weekend I adapted their code to autonomously discover the best way to improve a machine translation model for a low-resource language.

Man it feels good to have all the machines cooperating Friday night so they can train all weekend.

Just installed Claude Code

And about 2,000 languages not spoken by AI. We’re going to crush it.

In 10 years the internet will be an irrelevant source of data for training AI.

I’m surprised live tv still exists.