Essays on AI, robotics, inference optimization, and the future of intelligent systems.
Small neural networks rediscover the classical laws of the Riemann zeros, and a fake zeta function shows which laws depend on the Euler product. One pattern looked like a new law, until a literature check showed it was already a theorem. No new result, but a clearer map of what a proof must use.
Can tiny neural networks be reliable robot skills that chain into longer tasks? Two ~5,400-parameter skills on a simulated SO-101 arm, measured on 400 four-step tasks.
Most technology leaps come from three moves: make something static dynamic, add a dimension, or apply the system to itself. A simplified take on TRIZ, mapped to real innovations and to where LLMs go next.
What is TypeSafe's Jev model? A System One decision model that returns typed probabilities instead of text. How it compares to LLMs, and 7 use cases.
GPT-6 Astra's BenchCAD score and viral 3D demos are the visible part. The less-confirmed, more consequential claim: the same general-purpose model may already pick up an object it's never seen and place it, zero-shot, on a real robot arm — the same shift GPT-3.5 made for language, now happening for physical work.
Intelligence itself now has a price — cost per token — and it is falling toward the cost of human labor. A field guide to positioning for the moment the two prices cross.
The broad, freely accessible AI we use today is the opening move, not the endpoint. Here's how controllable AI will reshape who holds power over knowledge.
Fourteen training runs, $110 in GPU costs, and one working model later — what fine-tuning Gemma 4 on web navigation actually looks like.
A personal experiment log testing multiple compression techniques across two model scales, ending with a GPT-2 Large that runs 1.6× faster and outperforms its uncompressed baseline by 13.7%.
The biggest surprise of the AI boom may not be what AI can do — but how much it costs to run it at scale.
We've spent years building better retrieval. The problem isn't retrieval — it's the index. And generative AI finally gives us the tools to fix it.
A framework for understanding intelligence across its three forms — embodied, linguistic, and abstractive — and what it means that AI learned from the most complete one.
A personal essay on growing up obsessed with discovery, the longing to be at the frontier, and what happens when you finally get there.
We build harnesses around AI agents. A recent DeepMind paper asks whether the agent could have built a better structure than we did — and what it means that the answer is yes.
What happens when you apply the core ideas of Reinforcement Learning — rewards, exploration, delayed feedback — to your own life and career choices.
Why inference gets slower as context grows, and the emerging solutions to break through the linear attention barrier.
Not the hype, not the potential—where is real AI value being created right now?