What we're reading this week

Thirty months of AI progress in charts, including a GPQA correct answer falling from 30 cents to 0.04 cents. Mollick adds a third way of working beyond centaur and cyborg, the "conjurer" who summons agents and judges what comes back, and argues that even where AI is excellent, small gaps create large needs for people.

An Anthropic interpretability researcher on how models form concepts and plan ahead, and where that differs from surface prediction. A podcast, with no written essay.

Narrative conventions for showing minds (soliloquy, interior monologue, helped along by print and private reading) train readers' ability to model other people, passed on culturally like reading or chess. The evidence includes the Nicaraguan Sign Language case.

A first-person essay on why music changes how the world looks, weighing reductionist, cultural and metaphysical explanations and settling on emergence. The footnotes on audience connoisseurship and music as "axiom systems" are worth the scroll.

A Claude Code, FFmpeg and Whisper workflow where a TSV cut list is the whole edit, and anyone can inspect it. The cutting is scriptable and the comedic timing isn't: Brooke Lacey could direct the system only because she already had years of short-form experience.

How do you check that an LLM's code translation kept the behavior when the program is nondeterministic? One divergence doesn't prove a difference, and "different" gets murky once output sets, ordering and relative rates come in.

Maps 29 agentic benchmarks onto 923 occupations and finds the coverage skewed toward high-wage work. Mercor sells evaluation data, so it has a stake in the gap, but the map is a useful look at what AI progress is measured against.

An experiment with 50,000 made-up statistics, about $50 and seven months, showing how AI answers turn unverified numbers into apparent authority. The first half carries the finding; the rest is a how-to and a product pitch.

A 331-developer Microsoft and Oregon State survey plus DX's data from more than 500 organizations. It names the "hollow orchestrator," whose output grows while their ability to judge it may not, and is honest that the link is still a hypothesis.

Yena Lee, Britton Russell and Ben Hofferber from our team on the PHP flaws three teams with AI tools found in August, two years after the first signs were shelved. Finding bugs got cheap; patching still runs on someone else's schedule. The piece separates the systems a company can fix from the ones it can only ask its SaaS vendors about, and names three decisions to make before the next week like this one.

Imprint's numbers: four times the PRs per developer, median review time from 33 to 77 minutes, incidents down 42%. Larson argues verification is now the bottleneck, that passing agent output along unread makes it worse, and describes a CI-gated "small-change" path.

A seven-layer review pipeline: a classifier sets the risk tier and how many human approvals a change needs, verification agents vote on findings, a REVIEW.md file says what the repo cares about, past incidents become a daily eval, and after three rounds the reviewer asks the author to rethink the approach.

Anthropic built its own source control (Scion) and review tool (Antique) because the infrastructure, not the coding, became the bottleneck: 2,500 to 30,000 merges a day in four months, 10 to 15 agents per engineer, a team of seven. Includes the Mononoke "rip-off" argument and the Google and Meta precedents.

AI spend and PR throughput across 27 companies: company size doesn't predict spend, costs are falling with Opus 5.5 and open models, a two-developer shop does "the work of 5-7 people," and the tension between velocity and slop runs through all of it.

A marketplace does four jobs: search, transaction, making trust between strangers cheap, and fulfillment. Agents collapse the first two, so the businesses at real risk are the ones that do nothing else, with hotels as the main example.

Apple's HomeView and trust as a moat. The parts worth reading are the aside on using AI to decompile and rebuild your own software, and the point that committing to Siri AI means giving up every other AI.

Meta's new open model and a surprise Apple partnership to run it on the device, and what that means for where value sits as models move local. The full Update is paywalled.

Beck imagines a RoboGym: press "Lift" and watch a weight go up and down while a counter tallies the energy used. His worry is that AI tools are being built the same way for thinking. The rest of the post is for subscribers.

A designer with decades in works out how to use generative AI without losing his skills. He proposes treating it as an instrument, rather than a tool, a crutch or an ideology, and questions whether taste is really what sets people apart.

Psychological safety started as a research-backed condition for learning and drifted into "never feel bad at work." Majors goes back to Project Aristotle and eudaimonia to pull it back.

A build log of a small agentic setup with the rough parts left in: three cloud coding agents, a PM bot that SSHes into the Codex and Claude Code machines, a credential vault per agent, and Lean proofs gating changes. It ends on choosing big bets over a stream of small tweaks.

Will Larson on AI pulling technical decisions back to managers: hands-on leadership returning, hiring as the last resort, wider spans, and productivity without flow. The written takeaways stand on their own; there is also a version at lethain.com/revised-rules-of-engineering-leadership.

The case for sites exposing typed tools to agents instead of letting them click through the page from screenshots, with benchmark numbers and the standards fight behind it. The interface a site exposes decides what an agent can do there.