Notes on compute, cost, and the math underneath.
How we think about the training loop, the economics that make it matter, and the research the work is built on. No roadmaps, no secrets, just the ideas that keep us up at night.
The compute wall is the real ceiling of AI
The best ideas in AI are no longer the scarce resource. Compute is. A look at why the training bill, not the research, now decides what gets built.
4 min read →Everyone optimizes the machine. Almost nobody re-poses the problem.
Faster chips, tighter kernels, lower precision. The whole industry races to run the same computation faster. There is another axis, and it is wide open.
5 min read →Why reinforcement learning makes efficiency compound
In RL post-training, the model runs hundreds of rollouts for every learning step. Any per-run saving is paid out again and again. The math is unforgiving, and it cuts both ways.
4 min read →From energy-based models to modern training
What a decade of peer-reviewed research on quantum-assisted learning actually established, and the honest distance between that result and a frontier training run.
6 min read →