A Dollar Fifty in Late Charges

THE NUMBER: 96 – 48 – 12. Ninety-six: the class average on a Brown University take-home midterm this spring, with AI in the room. Forty-eight: the average when the professor moved the final in-person without warning, the lowest in the course’s history. Twelve: the lessons in Anthropic’s AI Fluency course, the closest thing yet to an instruction manual for the most powerful product ever shipped without one. All three are multiples of twelve, and the gap between the first two is the price of skipping the third. Hold those numbers. The whole issue hangs off them.

Boston, 1997. The bar scene in Good Will Hunting is remembered as a takedown, but watch what’s actually being priced. A Harvard grad student with a ponytail is running borrowed opinions on Chuckie, quoting a book he’ll disown by his second semester of grad school, using the material the way a mugger uses a crowbar. Will steps in, names the exact page the kid is plagiarizing his personality from, and then delivers the valuation: you dropped a hundred and fifty grand on “an education you coulda got for a dollar fifty in late charges” at the public library.

The scene works because everyone in it knows Will is right and nobody’s behavior changes. The ponytail kid still gets the Harvard degree, still gets the fund job. The market for credentials and the market for actual learning had decoupled, and in 1997 the arbitrage was invisible because the library card took years of solitary discipline to use and the diploma took a checkbook.

This week the arbitrage went visible. The library moved into your pocket, learned to talk, and started answering questions at 2 a.m. for twenty bucks a month. A professor at Brown accidentally published the first honest mark-to-market of what the credential measures now. And Anthropic, on Tuesday, did the thing we said on our Future Proof podcast Monday that nobody had done: started writing the user manual.

🎓 The Midterm That Marked the Diploma to Market

Start with the data, because it’s the cleanest natural experiment education has produced in years. Roberto Serrano has taught economics at Brown for 34 years. This spring he gave a take-home midterm. The class average was 96, against a historical range of 65 to 80. Forty of eighty-six students scored a perfect 100. Then he moved the final exam in-person, unannounced. Twenty-seven students didn’t show, twenty-two of them perfect scorers. Nineteen of the fifty-nine who sat for it failed. Some turned in blank exams with their names signed at the top. The average landed at 48. Paul Graham posted the scatter plot with the caption that it looks like all but three cheated. The provost, per the professor, never responded; a faculty committee called it “a wake-up call,” which is academia for we will form a subcommittee.

Everybody read that story as a cheating scandal. Jim Bianco read it better: these same students will sit at graduation in May and be told AI is the future, by the same institution that just flunked them for using it. A university’s actual job, on the Bianco reading, is producing the most AI-prepared workforce in the world, the way it produced the most internet-prepared workforce 25 years ago. Instead it’s producing muddled signals and blue books. The University of Chicago Law School just issued a new AI policy; Steven Sinofsky dug up the 1980s panic that word processors would destroy critical thinking. We’ve seen this movie before, except this time both sides are right. The students are gaming the assessment, and the assessment is measuring a world that no longer exists.

Here’s the uncomfortable synthesis: the take-home midterm and the in-person final weren’t measuring the same skill, and neither one measured the skill that matters. The midterm measured whether you can operate the machine. The blue book measured whether you memorized the textbook. Nobody measured whether you can tell when the machine is wrong, which is the only part an employer will pay for in 2031.

The bottom line for executives: Brown just repriced the transcript for you. A 4.0 now carries genuine uncertainty about which entity earned it. The fix costs one interview stage: a supervised work sample, tools allowed and disclosed. You’re not screening AI out. You’re screening for the kid who runs it versus the kid hiding behind it.

📚 Two Products, One Admission

Anthropic’s Tuesday launch got read as an education story. Read it as a strategy filing instead, because there are two products in the box and the quiet one is the interesting one.

The loud one is Claude for Teachers: free premium Claude, including Claude Code and Cowork, for every verified K-12 educator in the US who signs up before June 30, 2027. Not a stripped-down education SKU. The real thing, wired into something called Learning Commons, which gives Claude the academic standards for all 50 states plus the smaller competencies under each one and the order students typically learn them. It pulls from actual curricula (OpenSciEd, Illustrative Mathematics), builds differentiation plans for kids at different readiness levels, scaffolds for multilingual learners, accommodations for IEPs, and can take a student’s assessment history and draft an individualized plan overnight. Nine edtech integrations on day one. The American Federation of Teachers co-designed the privacy standard, Randi Weingarten supplied the blessing quote, the Gates Foundation is involved, and Detroit’s public schools pilot it in the fall. FERPA-compliant, no training on conversations.

Now notice the admission underneath, which analyst Poonam Soni caught within hours: student-facing AI has mixed results, and everyone knows it. So while every other lab builds the AI students use, Anthropic built the one teachers can’t work without. Those are very different companies in five years. Another sharp read, from the account SightBringer, called it what it is: a land grab for cognitive infrastructure. Students are fragmented users; teachers are force multipliers. Roughly three million of them, each touching 25 to 150 kids a year, forming the habits those kids will carry for a decade. The kids were never the customer. The people who program the kids are.

The union math makes the point sharper. The AFT’s own National Academy for AI Instruction, the $23 million deal it cut with Microsoft, OpenAI, and Anthropic last summer, aims to train 400,000 teachers over five years. Tuesday’s launch offered all three million free access in an afternoon. The union’s real moat was being the training bottleneck, and Anthropic just went around it, paying the toll in privacy standards and a press-release quote. Weingarten, meanwhile, is simultaneously on record wanting student-facing AI banned in elementary grades. Hold your nose or don’t; the deal got done.

And the quiet product: Claude for You carries a course called AI Fluency: Framework & Foundations. Twelve lessons, three to four hours, built with professors Joseph Feller (University College Cork) and Rick Dakan (Ringling College), organized around four skills with a rubric: Delegation, Description, Discernment, Diligence. What to hand off. How to specify it. How to judge what comes back. How to stay accountable for the result. On Monday’s podcast I said there has never been a more powerful product shipped to humanity all at once with literally no user manual. My partner Anthony pushed back that the manual is talking, in any language. Brown is the empirical rebuttal: those kids could talk to it fine. The midterm proved they’d mastered Delegation and Description. The final proved nobody had taught them Discernment or Diligence. Half a manual is how you get a 96 and a 48 from the same class.

Why this matters: the fluency course is the real deliverable, and the teacher product is its distribution vector. Standards mapping gets Claude adopted as the approved operating layer of instruction; the 4D habit is what actually transfers. Get them started early on your product and that is what they will use. Apple ran this play with classroom discounts in the eighties. Google ran it with Docs in the 2010s. Anthropic is running it with the teacher’s Sunday-night lesson plan.

🧭 The Insurgents Already Drew the Map

Don’t expect Claude for Teachers to move the national scores, though. AI teaching the same lessons to the same kids in the same rooms is a faster mimeograph. The system’s problem was never tooling. This year’s Census release put per-pupil spending at a record $17,619, up 6.6% in a year; the latest 12th-grade NAEP reading results, released in September, were the lowest ever recorded, with 45% of seniors below basic in math. More money, worse output, for decades, through every tool the tech industry ever donated. Alex Karp’s line cuts to the bone: all of our tests are built around things that were valuable in the Industrial Revolution. Not the AI revolution. Musk says the quiet part at full volume: colleges are a $200,000 receipt for proving you can do your chores.

The insurgents are aiming at the actual target, and the striking thing is what’s not inside their machines. The education-AI result of the year is a third grader scoring a 5 on AP Calculus BC through Math Academy, a system that contains no language model at all. Its engine is a knowledge graph of roughly 3,000 math topics whose prerequisite links one man, Justin Skycak, spent 250 hours encoding by hand, run through mastery gates and spaced repetition. It’s Bloom’s 1984 two-sigma tutoring result, finally industrialized. The Marble Curriculum just open-sourced its own version: 1,590 concepts, 3,221 connections across eight subjects, anchored to US and UK standards, a genuine DAG you can compute a learning path on. The intelligence was never the bottleneck. The map was. Which, you’ll notice, is exactly what Anthropic’s Learning Commons standards-mapping is: the map, productized, given away.

Then there’s Alpha School, the Austin operation expanding from 17 campuses toward 40 this fall. Two hours of adaptive academics a day, afternoons for workshops and life skills, adults recast as guides whose job is motivation and judgment rather than content delivery, paid a $100,000 floor (against a national teacher average of $74,495 and starting pay of $48,112) because the software runs instruction and the humans sell kids on wanting it. Be honest about the caveats: tuition runs $40K to $75K, the 2.6x learning claims are self-reported with no independent audit, and even the friendliest deep review concludes it works for maybe a third to two-thirds of kids. It is not the answer. It is Version 1.0 of the answer, at an early-adopter price, and the Brownsville campus already runs at $10K, which happens to be almost exactly the Texas voucher. Version 10, five years out, is 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.

What This Means For You

The thread through all of it: knowledge just got repriced to a dollar fifty, and everything scarce moved up a level, to desire and discernment. That’s true for your kids and it’s true for your payroll.

Take the manual, then issue it. Twelve lessons, three to four hours, free. Do it yourself first, then make the 4D course onboarding for every new hire. Every company is running an unmanualed power tool right now; yours can be the one that reads the card.

Reprice credentials in your hiring this quarter. Brown told you what a transcript measures now. One supervised work sample, tools allowed and disclosed, tells you what the transcript can’t: who runs the machine and who hides behind it.

Draw your firm’s knowledge graph before a vendor draws it for you. Skycak’s 250 hours of hand-encoded prerequisites beat every LLM tutor on the market, and Anthropic just demonstrated that the map, not the model, is what buys distribution. Your business runs on an unwritten map of which skills gate which outcomes. Write it down. That document is a moat no commodity model crosses.

At home, don’t wait for the district. The library in your kid’s pocket is open tonight. What the district won’t supply is the sales job: the desire to use it for something other than the take-home midterm, and the reflex to doubt what it says. That part was always the parent’s job. Now it’s the whole job.

Three Questions We Think You Should Be Asking Yourself

Could your team pass the in-person final on the work AI does for them? Brown’s gap was 48 points. Yours exists too; you just haven’t administered the exam. If a key process had to run for a day with the tools off, would you find operators or perfect-midterm no-shows?
Who’s writing your company’s manual? You bought the seats. Did you buy the fluency? If your AI training program is “here’s the login,” you’re running the same curriculum as the district: tool access, no discernment, and you’ll get the district’s results.
When the knowledge is free, what exactly are you paying for? In hiring, in tuition, in consulting: if the content costs a dollar fifty, the premium you pay should buy desire, judgment, and taste. Audit your biggest checks against that list and see which ones are really buying a ponytail quoting a book he hasn’t finished.

“There has probably never been a more powerful product shipped to humanity, all at once, with literally no user manual whatsoever.”
— Us, on Future Proof, Monday afternoon. By Tuesday morning, somebody had started writing it. Chapter one is twelve lessons. The rest gets written at your kitchen table.

— Harry and Anthony

Signal/Noise by CO/AI is published most weeknights from New Canaan, Connecticut. The point is to make you the smartest person in the room without taking more than fifteen minutes of your morning. If we did, forward it to one person. If we didn’t, hit reply and tell us why.