AI the News That's Fit to Prompt

Andrej Karpathy left OpenAI two years ago to teach one-person AI courses out of Eureka Labs. This week he joined Anthropic. At the same moment, Epoch AI quietly published a chart showing that high-bandwidth memory now eats 63% of the dollar a buyer spends on an AI chip's components, up from roughly a third two years ago. Both stories are about the same shift: the scarce thing is no longer the model. It is the talent that knows how to push it forward, and the memory that lets it run.

The Talent Layer

• Karpathy joining Anthropic is the headline, but the underlying signal is talent gravity. A researcher whose endorsements move developer attention has chosen the lab whose research culture and publishing posture he can live with, over a former employer that still leads on raw product velocity. Anthropic now has a magnet for the next round of recruiting.
• Wes Roth's read is that this is about interpretability and training pedagogy. Karpathy's last two years of essays gesture at exactly the agenda Anthropic publishes most aggressively on. The hire is less a surprise than a confirmation of where the most opinionated individual contributors are now choosing to land.
• OpenAI loses on perception even when it wins on product cadence. The pattern of senior researchers exiting toward Anthropic and a handful of independent labs is now a multi-year trend, not a quarter's worth of noise.

The Memory Layer

• Epoch AI's component-cost chart is the most underappreciated AI data release of the month. HBM rose from a minority share of AI chip BOM in Q1 2024 to roughly 63% in Q4 2025. Logic die costs barely moved; packaging shrank as a share of the whole. The next AI bottleneck has a name and a SKU, and it is high-bandwidth memory.
• That single chart reframes the capex story. The bottleneck is not GPUs or fabs. It is HBM allocation contracts and the handful of vendors who can produce it. The hyperscalers that locked in long-dated HBM supply in 2024 look different in 2026 than the ones that did not.
• Nate B. Jones's "boom about to hit a wall" video lands the same week. His thesis: power, fab capacity, memory supply, and enterprise willingness to absorb six-figure agent pilots are converging into a constraint set that current valuation models do not yet price. The Epoch chart is one of the data points behind that argument.

Open-Source Agents Keep Shipping

• Simon Willison shipped Datasette 1.0a30 plus Datasette Agent 0.1a4 on the same day. The base release adds a customizable "Jump to" menu and a plugin hook; the agent uses that exact hook to register a "Start a new agent chat" entry point. A hosted instance at agent.datasette.io is live with GitHub auth.
• Reasonix landed as a DeepSeek-native terminal coding agent. The pitch is a Claude-Code-shaped workflow at a fraction of the per-token cost. Another data point that the open-weights stack is reaching parity on the agent harness as well as the model.
• And then arXiv 2605.06445 documented the real ceiling. "Constraint decay" — LLM coding agents lose ~30 percentage points in assertion pass rates as structural requirements accumulate. The paper is the rare quantitative confirmation that agents do well on toy tasks and fold on real backends, which is what every team running an agent pilot has been saying out loud.

The Maintainer Pushback

• Armin Ronacher, via Simon Willison, names what every open-source maintainer is now feeling: long, plausible, and useless AI-generated bug reports are degrading the quality of issue trackers. Ronacher's ask is a return to short, specific, human-authored reports with the actual observation and error attached.
• Pair that with TechCrunch's piece on Google's AI security posture. Google Cloud's COO is publicly preaching security-from-the-start while Google itself is fielding billing-surprise and credential-revocation complaints. Both stories share a subtext: the AI tooling has shipped faster than the operational and social contracts around it.
• Amazon's Bee wearable is the consumer-side version of the same gap. Genuinely useful for professionals, immediately uncomfortable as a privacy proposition. The product is shipping; the norms for living next to it are not.

For the Practitioner

• Tina Huang's updated 2026 skills list has shifted away from prompting basics toward agent orchestration, evaluation discipline, and the production skills that distinguish a working AI engineer from a power user. The 2024 version of this video would have been mostly about ChatGPT tricks; the 2026 version is mostly about harness craft.

The Throughline

Today's stories cluster around one observation: the scarce inputs to the AI economy have changed. Two years ago the scarce things were model weights and the GPUs to train them on. This week's news names a different scarcity stack — talent that can move the frontier, HBM that can run it, enterprise budgets that can absorb the pilots, and policy frameworks that can hold the deployments. Karpathy's move and the Epoch chart are the two cleanest data points, but every other story in the issue is downstream of the same shift.

The Datasette releases and Reasonix are the optimistic side of the ledger. Small teams (or solo developers) are shipping open-source agents that match the workflow of closed runtimes, against a backdrop where the constraint decay paper says even the closed runtimes break in predictable ways once real structural requirements appear. That gap — between the demo and the deployed system — is now itself a researchable artifact, which means it is also a fixable one. The team that lowers constraint decay by ten points wins the next round.

The pessimistic side is the maintainer-and-user pushback. Ronacher's complaint about AI-generated bug reports, the privacy ambivalence around Bee, and Google's own credential-revocation stumbles are all instances of a single pattern: the tooling outran the operating practice. The next twelve months of AI productization will look less like a capability race and more like an operational-maturity race, and the teams that win it will be the ones that take maintainer feedback as a signal, not a nuisance.

The Bigger Picture

For three years the AI industry has marketed itself as a story about model intelligence. The most informative news this week is about everything underneath the model: who builds it, what it runs on, who pays for the pilot, who maintains the open-source it gets pointed at, and who is liable when it gets something wrong. That set of questions does not photograph well in a launch keynote, which is part of why it is being underpriced.

The talent gravity story (Karpathy to Anthropic) and the HBM story (Epoch's 63%) are early-warning indicators of the same structural shift. The next round of competitive differentiation is going to come from who you employ and what you can plug into a rack, not from whose benchmark number moved 3% this month. That is good news for the labs that have been quietly building research culture and supply contracts. It is bad news for the labs that have been building demos.

And underneath both is the maintainer-and-citizen reaction. Ronacher's quote, the Bee privacy unease, and the constraint-decay paper are all signals that the industry's customers — both the technical ones and the human ones — are beginning to push back on the parts of the AI rollout that have been over-promised. The labs that hear the pushback as feedback will adjust faster than the ones that hear it as friction.

What to Watch

• Anthropic's research output in the next six months. If Karpathy lands well, expect a noticeable uptick in interpretability and training-pedagogy publications. That is the visible signal that the hire is paying off internally.
• HBM contract disclosures in Q3/Q4 earnings. The Epoch chart says HBM allocation is now the dominant input cost. Watch for hyperscalers naming their long-dated HBM supply terms, and for analysts pricing the disclosure gap between the ones that do and the ones that do not.
• Follow-up work on constraint decay. If a credible team publishes a harness or training change that materially lowers the 30-point gap, the agent-runtime market opens up. If nobody does, expect the "agents in production" narrative to slow.