SIGNAL / NOISE
The Manual Arrives. The Reader Can't Read.
An economics professor at Brown gave a take-home midterm this spring. The class averaged 96. Forty of eighty-six students scored a perfect 100. Then he moved the final in-person, unannounced. The average fell to 48, the lowest in the course's history. Twenty-two of the perfect scorers never showed up. A few turned in blank exams with their names signed at the top.
On our Future Proof podcast Monday (give it a listen!) I said there has never been a more powerful product shipped to humanity all at once with literally no user manual whatsoever. My partner Anthony pushed back: the manual is talking - in any language. Brown is the rebuttal. The kids could talk to it fine. Nobody taught them the other three-quarters of the job.
Tuesday, Anthropic started writing the manual. The headline product is Claude for Teachers: free premium access for every verified K-12 educator in the US, lesson plans mapped to standards in all 50 states, the teachers' union blessing the deal, Detroit piloting it in the fall. It dropped at 11 a.m. Tuesday, hours after every AI newsletter had gone to press, so this morning your inbox is likely full of the headline. But the headline isn't the story. The tell is the admission underneath: student-facing AI has mixed results, so while every other lab builds the AI students use, Anthropic built the one teachers can't work without. The quieter product actually matters more. Claude for You carries a 12-lesson AI Fluency course built around four skills: Delegation, Description, Discernment, Diligence. What to hand off, how to ask, how to judge the answer, how to own the result. That's the manual. The teachers are the distribution vector.
Look, I don't expect Claude for Teachers to move the scores. AI teaching the same lessons to the same kids in the same rooms is a faster mimeograph. We spend a record $17,619 per pupil and 12th-grade reading just posted the lowest number ever recorded. The system isn't under-tooled. It's aimed at a world that stopped existing. Alex Karp said it plainly: our tests are built around what was valuable in the Industrial Revolution not the AI Revolution.
The insurgents are aiming at the right target. The AI tutoring system that just put a third grader through AP Calculus BC contains no language model at all, just a 3,000-topic knowledge graph one man spent 250 hours drawing by hand. Alpha School pays its guides $100,000 because software runs instruction and the adults sell kids on wanting it. Version 1.0 costs $40K a year. Version 10, five years out, is going to be mostly AI and very likely close to free. The teacher bet wins the next five years. The learner bet wins the decade. The bet on teachers is a hedge - the kids are the future.
Which brings us back to the bar in Good Will Hunting. Will tells the Harvard grad student he wasted 150 grand on "an education you coulda got for a dollar fifty in late charges" at the public library. The library now lives in your pocket for twenty bucks a month, and it never closes. It lends everything except the two things that were always scarce: a deep desire to learn, and knowing when to doubt what you're told. Math has answers. Code compiles or it doesn't. History, art, strategy, interpersonal skills — those take judgment, and no district curriculum on earth teaches it.
The knowledge now costs a buck fifty. Desire and doubt are the tuition.
At COAI today: the full Signal/Noise, with the Brown data, the Alpha economics, and the two-product read on Anthropic's launch, is live at getcoai.com.
Whether your team can delegate, describe, discern, and stay accountable — that's the conversation we want to have. If Tuesday's launch put AI fluency on your calendar, let's chat.
ONE — A NUMBER THAT SUMMARIZES THE DAY
96 – 48 – 12. Ninety-six: the average on Brown's take-home midterm, with AI. Forty-eight: the average when the final moved in-person, without it. Twelve: the lessons in Anthropic's new AI Fluency course, the first real instruction manual for the most powerful product ever shipped without one. All multiples of twelve, and the gap between the first two numbers is the price of skipping the third.
THREE — ACTIONS TO TAKE TODAY
1. Take the manual yourself, then make it the default. Anthropic's AI Fluency course is 12 lessons, three to four hours, free at anthropic.com/learn. Delegation, Description, Discernment, Diligence. Block the time today and assign it as onboarding for every new hire. You already pay for the tool; this is the missing operator's card.
2. Reprice the diploma in your hiring. Brown just marked the transcript to market: 40 perfect midterms, a 48 average when the laptops closed. Add one supervised work sample to your interview loop today, done live, tools allowed and disclosed. You're not screening AI out. You're screening for who can use it and who's hiding behind it.
3. Start your company's knowledge graph. One tutor spent 250 hours mapping 3,000 math topics and beat every LLM on the market. Your firm runs on an unwritten map of what depends on what. Open a doc today and have your best operator sketch the ten skills that gate everything else. Models are commodities. The map is the moat.
FIVE — STORIES TO KEEP YOU INFORMED
Wednesday, July 15
Anthropic enlists America's teachers. Free premium Claude for verified K-12 educators through June 2027, standards-mapped in all 50 states, AFT partnership, Detroit pilot. The kids were never the customer; the three million people who program the kids are. (Full analysis above.)
16 Nobel laureates put the job shock on the clock. The "We Must Act Now" letter, 200+ economists including OpenAI's and Anthropic's own, warns AI gives us "only a few years" to adapt. AI-cited US job cuts: 101,743 through June, nearly double all of 2025. The petition asks for institutions. See above for what shipping one looks like.
The rivals co-sign the referee. Hassabis proposed a US-led frontier-AI watchdog with slowdown powers; Altman and Nadella both endorsed it within hours Tuesday. When the racers ask for a pace car, ask who owns the pace car.
Anthropic extends Fable 5 a third time as Kimi K3 looms. Free frontier access through July 19, Claude Code limits up 50%, with Moonshot's next model rumored inside 24 hours. Four frontier models in one month and no lead survives two weeks. The treadmill is the product now.
SK Hynix says demand outruns supply until 2030. The stock dropped 20% in two days. Korean regulators counted 344 billion won in forced liquidations in nine July days as leveraged-ETF holders blew up. The volatility is in the paper, not the demand. Position accordingly.
— Harry and Anthony
Sources: