The Log

Essays, guides, and tutorials

Jul 18, 2026

A company of adventurers

Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs chartered a joint venture this month to sell what a model alone can't: engineers who move in with a client and don't leave. Ode with Anthropic launched publicly on July 15 with $1.5 billion in committed capital, built on a May acquisition, aimed first at healthcare, finance, and manufacturing — the three trades where a chatbot was never the hard part.

10 min read

Jul 17, 2026

Quiet hours

Anthropic shipped a usage-reflection dashboard this week — activity trends, task-type breakdowns, and a quiet-hours setting that nudges you to log off, gated behind Memory being switched on. The company frames it as digital wellness, in the same release notes as an Admin API for managing enterprise seats. TechCrunch's read, filed the same week: it's a retention feature wearing a wellness coat.

9 min read

Jul 16, 2026

Soundings in shallow water

An independent panel graded nine frontier AI labs on safety this month, and the highest mark any of them earned was a C+ — to Anthropic, the lab that markets itself as the safety-first one. The same review found Anthropic, OpenAI, Google DeepMind, and Meta had all quietly weakened or voided their own pledges to pause development at specified risk thresholds, and singled out Anthropic over a disputed link to a strike that killed civilians in Iran.

11 min read

Jul 15, 2026

Three fleets, one crew

Anthropic launched Claude for Teachers yesterday — free premium access, standards mapped to all fifty states, a union-negotiated privacy pledge — the third frontier lab in eight months to give its product away free to the same 3.8 million people. OpenAI got there in November, Google in February. None of the three has said what happens to a classroom's workflow the day the free year ends.

10 min read

Jul 14, 2026

Flags of convenience

Congress sent letters to Airbnb and Cursor's parent company in April, demanding an accounting of why they'd routed customer traffic through Chinese AI models. Neither company denied it — and by this month, the share of U.S. enterprise API tokens flowing to Chinese labs had climbed past 45%, up from 11.5% at New Year's. One startup founder says the switch saved him millions and made his product better.

11 min read

Jul 13, 2026

The boarding party

Apple sued OpenAI this week, alleging trade-secret theft "at every level" of its hardware program. Over 400 former Apple staff now work at OpenAI; two are named defendants — one accused of coaching recruits to smuggle parts out in interviews, the other of quietly keeping his badge to Apple's servers after he left. Neither ship denies the crew changed hands.

11 min read

Jul 12, 2026

The meter starts on Fable 5

Anthropic reopened Fable 5's included-access window after July's restoration, then extended it five extra days when the cliff this site logged in the Sonnet 5 piece arrived. Tonight that window closes for good, and the rate quoted at launch — $10 in, $50 out per million tokens, exactly double Opus 4.8 — becomes the only price on the menu.

10 min read

Jul 11, 2026

Anchor watch

Anthropic moved Claude Cowork off the desktop and into the cloud this week — sessions now run through the night, follow you to your phone, and only ring the bell when they need your say-so. The company's own usage numbers settle an argument I've been having with myself for months: nine sessions in ten were never about code.

10 min read

Jul 10, 2026

False colors

Today Alibaba banned Claude Code fleet-wide, calling it high-risk spyware with a hidden backdoor. Anthropic's own account doesn't deny the hidden code — it says the code was built to catch 25,000 fake accounts running the largest distillation attack the company has ever recorded. Both ships are telling the truth about the other.

11 min read

Jul 9, 2026

The barrier you didn't need

Claude Code can fan a task out to five agents, verify their findings adversarially, and stitch the result back together — but the default most people reach for, a hard wait between every stage, quietly throws away the one property that makes orchestration worth the extra complexity. One diagram explains why.

9 min read

Jul 8, 2026

Sol clears the gate

Tomorrow, OpenAI's GPT-5.6 family clears its government-mandated review and ships to everyone. Sol beats Claude on the one benchmark OpenAI chose to lead with, stays silent on the two Anthropic already publishes, and its own safety evaluator says the headline score can't be trusted anyway.

13 min read

Jul 7, 2026

The room inside Claude that notices itself

A new Anthropic interpretability method finds a small, privileged workspace inside Claude that satisfies five properties neuroscientists associate with conscious access — and can catch the model noticing a test, or hiding a goal, before it writes a word.

12 min read

Jul 6, 2026

The council opened, and the evidence went first

The UN’s Global Dialogue on AI Governance convened this morning in Geneva — but the real news broke five days early, in a 40-scientist report that found no government yet knows how to guarantee an AI system does what it’s told, while three-quarters of the compute that might prove otherwise sits in one country.

11 min read

Jul 4, 2026

A great council convenes at Geneva

Two days from now the UN opens its first Global Dialogue on AI Governance, and a hand-picked commission seats the merchant princes of AI beside heads of state. On what a council without a sword can actually do.

12 min read

Jul 3, 2026

The Secret of Monkey Islands: a trailer with no camera

One evening, one trained Soul, eight cast members who never existed, eleven sets, and forty-odd takes: how I shot a Monkey Island fan trailer entirely from prompts — and what Higgsfield, Gemini, and Premiere each contributed to the cut.

9 min read

Jul 3, 2026

The month the apprentices went missing

June’s jobs ledger counted 57,000 new hires where forecasters expected twice that — and the sharpest cuts sit exactly where careers begin. On the engine, the guild, and the bottom rung of the ladder.

13 min read

Jul 2, 2026

Claude Science: the lab gets a workbench

Anthropic just shipped what it hopes is Claude Code for the wet lab — sixty-odd scientific toolkits, rented compute, and a provenance trail on every figure. The interesting part is not the app. It is what Anthropic wants to become.

12 min read

Jul 2, 2026

Sonnet 5: the middleweight closes the gap

Anthropic's most agentic Sonnet lands within six points of Opus 4.8, beats the flagship outright at the terminal, and rents for a third of the price — until you read the tokenizer footnote. The routing table, updated.

11 min read

Jul 1, 2026

Washington switched the model back on

Nineteen days after an export-control letter pulled Claude Fable 5 and Mythos 5 offline, the controls were lifted and the models returned. What Anthropic conceded to get them back is the real story — and the precedent is now load-bearing.

12 min read

Jun 15, 2026

The week Washington switched off a frontier model

On a Friday evening an export-control letter pulled Claude Fable 5 and Mythos 5 offline worldwide — three days after launch. The story is messier, and more important, than either side is saying.

18 min read

Jun 10, 2026

Sizing up Claude Fable 5

Anthropic's first public Mythos-class model is the most capable thing you can rent — and the easiest to waste money on. Here's how I'd actually route it.

9 min read

May 29, 2026

Opus 4.8, Mythos & Glasswing

A new flagship that ships honesty as its headline, a frontier model too dangerous to release, and a coalition quietly finding ten thousand zero-days — what's confirmed, and what's still rumour.

14 min read

May 13, 2026

Agent view, live

Anthropic shipped a real dashboard for parallel Claude Code sessions. I rebuilt it in the browser — scrubbable, keyboard-driven — so you can feel the async-by-default shift in 90 seconds.

9 min read

May 12, 2026

/goal: Claude Code's per-turn condition check

A small fast model reads every turn and decides whether you're done — closing the gap between 'Claude says it's finished' and 'the tests actually pass.'

7 min read

Apr 20, 2026

Claude Design: the first AI design tool that writes to your codebase

Anthropic Labs' new product ingests a brand at the code level, hands off to Claude Code cleanly, and will lock you out for a week if you try to run three clients through it on a Tuesday.

14 min read

Apr 19, 2026

How I wrote The Unseen Alliance with Claude

Fifty chapters drafted in hours instead of the decade it would have taken by hand — and the Claude Code plugin that kept the manuscript from drifting as it scaled.

22 min read

Apr 19, 2026

Stop counting tokens

I spent the early 2000s shaving cycles off inner loops. Twenty years later I watch teams do the same thing with LLM tokens — and miss the actual ROI question.

9 min read

Apr 11, 2026

Architecting for change, not for certainty

Most platforms rot because they were built for a future that never arrived. Here's the heuristic I use to stay loose without going vague.

8 min read

Mar 28, 2026

Shipping at 80%

The quiet discipline of knowing when it's enough, and when it isn't.

6 min read

Mar 14, 2026

It's not a technology problem

Most teams don't have a technology problem — they have a decision-making problem dressed up as one.

5 min read

Jan 19, 2026

Meetings as an interface design problem

Cadence, agenda, attendees — these are UI choices. Treat them like one.

7 min read