Applied AI reviews, build logs and essays on systems, validation, agent workflows, decision-making and how ideas survive contact with reality.
What if every mistake made by one AI agent automatically made every other agent smarter tomorrow? One of the strangest things about modern AI agents is that they have terrible institutional memory. A human employee learn
Read essay ->
A new paper argues that LeJEPA can recover the hidden variables of the world from pixels, but only under one very specific condition: the world had better be Gaussian. There is a beautiful and slightly dangerous dream be
Read ->
For years, unstructured sparsity in neural networks was the AI equivalent of a concept car: beautiful in theory, useless on real roads. This paper may have changed that. There is a deeply irritating problem in modern AI.
Read ->
Anthropic’s secretive Claude Mythos rollout may say less about AI apocalypse and more about the brutal economics of frontier inference. Anthropic’s new Claude Mythos Preview has acquired the sort of mystique usually rese
Read ->
A 100-page survey argues that code is no longer just what agents produce. It is becoming the executable, inspectable, stateful infrastructure they live inside. There is a slightly ridiculous fantasy in AI right now. It g
Read ->
Bjarne Stroustrup is right to worry about vibe coding. The problem is not that AI writes code. The problem is that it lets people create technical debt at industrial speed. Bjarne Stroustrup, the creator of C++, recently
Read ->
A Princeton-led paper shows how to train VLM agents for 100+ turn decision-making without using another giant model as the critic. The trick is almost rude in its simplicity: let the big model act, and let a tiny CNN jud
Read ->Karpathy’s latest conversation makes the shift obvious: AI is no longer just helping people write code. It is changing what software work…
Speculative decoding was already clever. DFlash makes the draft stage parallel, turning diffusion from a clumsy text generator into a very useful accelerator. Large language models have a very expensive habit. They think
Ed Zitron asked a useful question for once. His conclusion is too theatrical, but the mystery is real: Microsoft’s AI buildout is much…
