This strange square 👇 is undoubtedly the most extraordinary work of literature in human history. Yet, unfortunately, barely anyone in the West has ever heard of it. There was this woman poet in 4th century China called Su Hui (蘇蕙), a child genius who had reportedly mastered

My new blog post discusses the physical reality of computation and why this means we will not see AGI or any meaningful superintelligence:

Seven CEOs and one scientist who's only half in the picture

I had Nano Banana remove GPT5.2's bounding boxes and Gemini 3 give it a go Left: GPT5.2 Right: Gemini 3.0

One hour Waymo ride from San Francisco to Menlo Park via highway 280. It’s over, cars are self-driving. Everything else is just about rolling this out to the rest of the world.

A year ago, we verified a preview of an unreleased version of @OpenAI o3 (High) that scored 88% on ARC-AGI-1 at est. $4.5k/task Today, we’ve verified a new GPT-5.2 Pro (X-High) SOTA score of 90.5% at $11.64/task This represents a ~390X efficiency improvement in one year

Larry Page & Sergey Brin had the PageRank paper (the algorithm behind Google Search) rejected. A reviewer called it “disjointed.” Geoffrey Hinton's Dropout was rejected for being “too simple.” I often feel the academic peer review is like a random process, especially when a

"Modified MIT" license: "You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million…"

(Fun fact: the distillation paper was rejected from NeurIPS 2014 because it was "unlikely to have significant impact").

Wow cool! New research on, literally, "why ice is slippery". :D

There are competing views on whether RL can genuinely improve base model's performance (e.g., pass@128). The answer is both yes and no, largely depending on the interplay between pre-training, mid-training, and RL. We trained a few hundreds of GPT-2 scale LMs on synthetic

@_arohan_ I know what you mean, but it's really frustrating because Google would NEVER have allowed any of it if they didn't come under very, very intense competition pressure. I remember explicitly choosing to avoid image generation topic in our team (despite being excited and having

The people saying AI can't generate "new knowledge" come off as pretty crazy to me. I can generate new knowledge with a few old fashioned Google searches. Weird overconfident ontological statement that doesn't hold up.

I do think many people over-claim on the level to which LLMs can do interesting new things. But to say they "cannot generate knowledge" is simply wrong. Like humans, they can interact with the world, and generate knowledge in the process.

This is incorrect. LLMs can call tools to get info and change about the outside world, including viewing videos, moving robotic arms, etc. The pic below shows a minimal falsifying example. The value of this number squared is new information—it's never been documented before.

Wild that saying “maybe we shouldn’t normalise public executions” is being floated as a “far-left/insane” position by US tech figures. The West stopped doing public executions because we moved beyond barbarism. One would think a founder of a university would know this.

Oh no. It has gotten so bad that even Paul Krugman is opining on GPUs and TPUs. No Paul, the G in GPUs isn't for general.

And this is the easy bit. Tmux is another level again.

Please, if you're a terminal company and you have a meeting about releasing a video featuring a mouse, stand up and say something.

Here's the hide and seek maps, ready for searching out the words!