Diet

Stripe’s Cheeky Pint with Mark Zuckerberg and Alexandr Wang on Meta’s Muse Wonder — breakout use cases, the model as the product, how frontier models get built, and Meta Superintelligence Labs.

On Tesla FSD: same pattern as harness learning — the model swallowed more and more of the C++ until nothing was left.

On training a proposer to revise an agent’s harness from execution feedback — test-time adaptation without parameter updates, so the model can swallow more of the scaffolding over time.

On the widening gap in shared understanding of LLM capability — from ~6B people barely touched, through casual free-tier use, to a thin professional sliver seeing the curve go vertical.

Caltech 2026 slides on AI disrupting mathematical research — from proof scarcity (Math 1.0) to proof abundance, and what changes when solutions become cheap to generate and verify.

On why agent runtimes should invest in ephemeral sandboxes — state, credentials, isolation, and scale are solved problems if you put in the work, not reasons to default to persisted VMs.

On treating frontier models like insider risks — separate the supply of intelligence from authority over it, and surround non-deterministic models with deterministic controls, observability, and independent auditability.

On Neo dropping pro rata rights in priced equity rounds — follow-on allocation should be earned through the founder relationship, not locked in by contract.

An informal reflection on what is happening in AI, recorded for Boston College’s Investment Committee.

Coatue’s October 2026 public-markets update.

Anthropic’s alignment science lead on existential risk — personally >10% within the next decade, with no clear plan yet for superintelligence alignment.