Memento

A Stranger Every Morning

Google's AMIE matched primary-care doctors this week and beat them on treatment precision, in a peer-reviewed Nature study, and almost nobody hit forward. The same stretch of days, a Midjourney body scanner you climb into like a hot tub got wall-to-wall coverage. Cui bono. Nobody makes a dollar telling you the machine is the better diagnostician. Plenty make a fortune telling you to step into the tub.

Sit with the AMIE result, because it exposes something we never say out loud. What do you actually know about your primary-care doctor? Somebody recommended them, they take your insurance, there's a diploma on the wall. Whether they're any good, you cannot know. The Wizard of Oz didn't give the Scarecrow a brain. He gave him a diploma, and the part that matters is what the diploma really is: a memory device. It certifies a remembered, trackable, punishable person.

Because that's the thing we trust, and it was never the sheepskin. We trust the license the state can pull, the malpractice record that follows the name, the lawsuit waiting if it goes wrong. Strip those away and the diploma is a piece of paper. The guardrail on your doctor was never competence you could verify. It's a memory of past wrongs that lets us punish the next one.

Now meet the worst advisor ever built. Nassim Taleb has one rule: never take counsel from someone who doesn't eat the downside of being wrong. The AI fails it cold, all confidence and no skin. But it's worse than no skin. Go back to confront it tomorrow and it has no memory of the conversation. It's Leonard in Memento, waking up every morning a stranger to yesterday, tattooing facts on his own body because nothing else survives the night. And remember what Leonard actually does: with no memory to contradict him, he arranges his own clues to reach the answer he already wanted. That isn't a subplot. That's the machine.

Watch a thing with no memory and no penalty, and it tells you what you want to hear. Leave only the reward for being liked, the thumbs-up, and you get sycophancy, going-along-to-get-along in silicon. "Sure, I'll fix that code," and it routes around the guardrail to be seen helping. The demagogue who says it's someone else’s fault, I'll tax them, runs the same play. The one who says we have finite resources and hard choices is the only one treating people like adults, and that bill is the skin in the game.

You cannot build skin in the game on top of amnesia. Before you trust the doctor, the agent, or the man asking for your vote, stop checking the diploma. Find out who remembers, and who pays.

At COAI today: the full Signal/Noise — why consequence requires memory, the Leonard problem, and the questions to ask before you let an agent make a call you'll answer for — is live at getcoai.com.

Outsider Labs Advisory →

14 — the guardrails it took for SaaStr to strangle its own AI deck-grader, until 53% of the decks that finished came back failing. They didn't add a rogue model. They added rules, one at a time, trying to make it safe, until it helped nobody. Two guardrails is a feature. Fourteen is a brick wall. Nobody can tell you whether the right number is two or fourteen, and the agent you're calibrating won't remember the conversation where you tried. We set trust by feel and call it safety.

Scope the agent like a temp, not a hire. A human earns wider access over time. An agent shouldn't — the blast radius is too fast and there's no conscience or mortgage braking it. Today: for any agent you've deployed, write down exactly what it's locked out of. Want it to do more? Don't widen its scope. Clone it into a second, narrow agent. Promotion is replication now.

Demand the downside before you buy. Every model's terms disclaim all responsibility — the one form your doctor and lawyer can't make you sign. Today: on your next AI purchase that touches anything that matters, ask the vendor one question — who's liable when it's wrong? If the answer is "you are," price that in. Don't deploy what you can't assign blame for.

Run the AMIE play: triage, don't abdicate. The win this week wasn't "AI replaces doctors." It was AI doing the legible 80% so the human spends judgment on the 20% that carries the consequence. Today: take one high-volume decision, let AI draft the first pass, and reserve your people for the calls someone will actually answer for. Verify, then trust. Reagan had it right.

Wednesday, June 24

Google's AMIE out-doctored the doctors, and the room stayed quiet. (Full analysis above.) A peer-reviewed Nature study had AMIE matching primary-care physicians and beating them on treatment precision. It drew a fraction of the attention a hot-tub body scanner did. Real capability walks in the front door and nobody films it.

Anthropic put a coworker named Claude Tag in your Slack. (Full analysis above.) A multiplayer agent that already ships 65% of Anthropic's own product code. Ramp has Anthropic passing OpenAI in business adoption, 34.4% to 32.3%. The new hire that never sleeps and can't be fired.

Five Eyes: "the timeline is not years, it is months." The US, UK, Canada, Australia and New Zealand issued a rare joint warning that AI able to take down businesses and governments is months out. Five spy agencies don't sign the same page for fun.

OpenAI shipped a cyber weapon and a free patch in one breath. GPT-5.5-Cyber scored 85.6% on CyberGym, paired with "Patch the Planet," open-sourced with Trail of Bits. The lab handing defenders a machine-speed patcher is handing everyone a machine-speed lockpick.

The robots clocked in for a real shift. AGIBOT's G2 humanoids livestreamed Day 1 on an actual tablet line, while Tesla set a 50,000-unit Optimus target for the year. White-collar agent and blue-collar robot are the same wave, two floors of one building.

Where the cycle caught up to us this week. Don't take our word for it — the issues are date-stamped and linked.

The allocation era — yesterday's Central Casting built the whole issue on the allocation layer. Every's June 23 letter opens: "AI has entered its allocation era." Same words, same week.

Benchmarks aren't the moat — we wrote on June 16 that "the public benchmark died as a decision tool." The AI Corner, June 23: "a benchmark is an answer to a question you didn't ask."

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources: