What it takes to make agents useful once generating code is the easy part. Eleven takeaways from Vercel Ship: durable execution, portable context, generative UI, and verification engineering—with diagrams and frames from the livestream.
Software stops being built and starts being summoned. The paradigm shift, the protocols (AG-UI, A2UI), the three structural patterns, PAGEN benchmark, vibe coding's QA crisis, and the bottlenecks that decide whether the whole thing scales.
How frontier AI is being made transparent (mechanistic interpretability) and adaptive (continuous learning). Superposition, sparse autoencoders, the pragmatic pivot, catastrophic forgetting, and what changes for working data scientists.
24 months of frontier-model evolution as a field guide. Knowledge cutoffs, MoE routing, the synthetic-data loop, the transparency collapse.
Carl Benedikt Frey's history of automation, translated for AI-era founders. Labor-enabling vs replacing, Engels' Pause repeating now, why the cluster traps itself, and TAM ≠ buyer.
Six months of featured launches (3,311 of them), regressed with confidence intervals.
Karpathy's 3.5-hour deep dive: tokenization quirks, SFT, RLHF, RL on reasoning, plus six pieces of "LLM psychology."
His 1-hour intro distilled: two files, the zip of the internet, the reversal curse, the LLM OS analogy.