The barrier just fell.
“The hottest new programming language is English.”— ANDREJ KARPATHY (CO-FOUNDER, OpenAI; EX-DIRECTOR OF AI, TESLA), 2023
For seventy years, using a computer to its full power meant learning to speak its language — code. That just ended. Now you say what you want, in plain English or Spanish, and the machine writes the program, runs the analysis, reads a thousand papers before lunch.
So the bottleneck moved. When anyone can generate anything, the scarce things become the ones a machine can’t hand you: a good question, and the judgment to tell a true answer from a merely convincing one.
Even Anthropic describes the human job, inside its own AI-written codebase, as building “the scaffolds… to let us trust it.” The scaffolds, and the trust. That is the whole skill now.
Generating is becoming free. Knowing what’s worth building — and checking that it’s actually right — is not.
So this school isn’t for engineers. It’s for anyone with good questions and the will to check the answers — and it aims these machines at the one thing only you have: your field, your data, your corner of the world.
In three years, these machines went from autocomplete to the frontier of science.
This is not a forecast. Every milestone below already happened, and every one is linked to its primary source. Watch the climb — then we’ll talk about what it does and doesn’t mean.
It started as a party trick.
ChatGPT arrives. It writes essays, emails, and code in seconds — and states false things with perfect confidence. Useful, dazzling, and not yet to be trusted.
OPENAI · NOV 2022 ↗A mother, 17 doctors, one chatbot.
After three years of unexplained pain and seventeen specialists, a mother typed her son’s symptoms and his own MRI notes into ChatGPT. It suggested a rare condition — tethered cord syndrome — that a neurosurgeon then confirmed and corrected in surgery. The AI spotted the pattern. Humans did the medicine.
TODAY · SEP 2023 ↗It solved a 50-year-old problem in biology.
AlphaFold predicts the 3D shape of nearly all 200 million known proteins — folding that had stumped chemists for half a century. Demis Hassabis and John Jumper win the 2024 Nobel Prize in Chemistry for it.
NOBELPRIZE.ORG · OCT 2024 ↗It won gold at the Math Olympiad.
Google DeepMind’s Gemini earned an officially graded gold medal at the International Mathematical Olympiad — 35 of 42 points, five of six problems, written in plain language under the same time limit as the students. The IMO president confirmed the score on the record.
DEEPMIND · JUL 2025 ↗It broke a record that stood since 1969.
An AI coding agent, AlphaEvolve, found a faster way to multiply 4×4 matrices than the algorithm humans had relied on since 1969 — and improved the best-known answer to roughly 20% of fifty-plus open math problems it was set on.
DEEPMIND · MAY 2025 ↗It disproved an 80-year-old conjecture.
An OpenAI model overturned a belief mathematicians had held since 1946 about a problem in geometry — and leading mathematicians checked the new result by hand. Nature’s headline: “AI cracks 80-year-old mathematics challenge — researchers are astonished.”
NATURE · MAY 2026 ↗The machine started writing the machine.
Anthropic reports that Claude now authors more than 80% of the code merged into its own codebase. The thing you saw at the top of this page builds the next version of itself — with humans directing and reviewing.
ANTHROPIC · JUN 2026 ↗Most of those were billion-dollar labs. One was a mother with a question and a chatbot. The distance between them is shrinking — and that is the opening.
Point it at what only you have.
None of this is abstract. By the end, you’re using these machines on real work — in whatever field is already yours. A few of the directions people take it:
Read the literature in days, not years.
Synthesize hundreds of papers, surface the question nobody has answered yet, design and pressure-test a hypothesis — then verify every citation so you don’t build on sand. It’s how this school’s sibling projects in chemistry and veterinary medicine actually run.
Find the whitespace, not the average.
Most people ask the machine “give me a business idea” and get the same answer as everyone else — the mean of the internet. You’ll learn the opposite: map a market, find the gap nobody is serving, and structure the idea until it’s real.
Build a real edge, not generic output.
From research to script to the video and design pipelines creators are quietly using — pointed at an angle only you can own, instead of the sameness everyone else is publishing.
An amplifier of what you already know.
A trader, a teacher, a vet, a chemist, a designer. The leverage was never the tool — it’s the tool aimed at your knowledge, your city, your data. That’s the part no one else can copy.
Everyone sells the first half. Almost nobody teaches the second.
We’ll absolutely teach you the tools from zero — yes, including basic Claude. But the reason to stay is the half that’s missing everywhere else: using these machines with judgment.
Use everything that exists.
Claude and Claude Code from zero. The plugins and skills almost nobody uses. MCP — plugging your AI into GitHub, your files, the web. Searching repositories like a researcher. RAG, embeddings, agents: every term, demystified once, properly, with a mental image that sticks.
Lead it — don’t follow it.
How to evaluate any tool in six questions. Why a model’s confidence tells you nothing about whether it’s right — and why it will agree with you even when you’re wrong. Quote-and-verify: no claim enters your work unchecked. The oracle principle: delegate exactly as much as you can verify.
Build the system that builds for you.
Memory, automation, pipelines that run while you sleep — with the audit loop that keeps them honest. Then point all of it at the thing only you have: your field, your data, your corner of the world. An amplifier of what you already know.
The oldest skill, suddenly the most useful.
The half we keep calling “judgment” has a real name. Epistemology — the study of knowledge: how we know what we know, and how to tell a justified belief from one that merely feels true. For most of 2,400 years it was a philosopher’s concern. This year it quietly became one of the most practical skills on earth. Four reasons why.
These machines are fluent and unreliable at once. Three philosophers argue their output is best understood as “indifferent to the truth” — stated with the same confidence whether it’s right or wrong. So a convincing answer is not evidence of a correct one. Telling the two apart is epistemology.
HICKS, HUMPHRIES & SLATER · ETHICS & INFO. TECH. · 2024 ↗Ask the question everyone asks and you get the answer everyone gets. Researchers find a model’s output is less epistemically diverse than a plain web search — a quiet collapse toward the average. Your edge is the question only you would think to ask, and the answer you actually check.
EPISTEMIC DIVERSITY & KNOWLEDGE COLLAPSE · arXiv · 2025 ↗For a great many tasks it’s far easier to check an answer than to produce one — OpenAI’s Jason Wei calls it the asymmetry of verification. That gap is your leverage: let the machine generate, and win by verifying well. Not always — fact-checking an essay can cost more than writing it. Knowing which is which is the skill.
JASON WEI · JUL 2025 ↗Mapping ~200 frontier researchers to build this school, our own work hit a wall the field rarely names: the model steers your questions toward what’s already common in its training — so its blind spots quietly become yours. The discipline is dragging those defaults into the light. That, too, is epistemology — turned on the tool itself.
FROM OUR META-RESEARCH ON AI PRACTITIONERS · 2026And here is what makes it permanent: the tools change every few weeks. The day we opened this school, Anthropic shipped Fable 5 — by its benchmarks, state-of-the-art on nearly every test it was given. By the time you read this, there is something newer. The specific tool is always passing. The ability to judge what it produces is the one thing that compounds instead of expiring.
That is the whole bet. Don’t chase the model of the month — build the judgment that outlasts every version of it.
This dream is seventy years old. It just shipped.
1956
W. Ross Ashby names the project
“…intellectual power, like physical power, can be amplified. Let no one say that it cannot be done… What is new is that we can now do it synthetically, consciously, deliberately.”
— An Introduction to Cybernetics, the closing paragraph of the book
1960
J.C.R. Licklider imagines the partnership
“…human brains and computing machines will be coupled together very tightly, and the resulting partnership will think as no human brain has ever thought.”
1962
Douglas Engelbart writes the framework
“By ‘augmenting human intellect’ we mean increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems.”
1970
John Conway proves the principle you’re watching
Three rules on a grid — born with 3 neighbors, survive with 2 or 3 — and complexity without limit emerges. Richard Guy spots the first glider (1969); Bill Gosper builds the gun that fires them forever (1970). The animation behind the top of this page is that exact mathematics, published by Martin Gardner in Scientific American, October 1970. Simple rules, amplified consequences.
1996
Fred Brooks states the inequality
“My thesis is that IA > AI: intelligence amplifying systems can, at any given level of available systems technology, beat AI systems. A machine and a mind can beat a mind-imitating machine working by itself.”
2003
The glider becomes the hacker emblem
Eric S. Raymond proposes Conway’s glider as the emblem of hacker culture: “simple rules of cooperation with what’s nearby lead to unexpected, even startling complexities.” It’s our mark too — worn with the verification habit it deserves.
2026
The amplifier starts building itself
Anthropic publishes “When AI builds itself”: most merged code at Anthropic is now written by Claude. The loop is closing — which makes the human role sharper, not smaller.
The dream was never artificial intelligence replacing you.
It was your intelligence, amplified. The amplifier has arrived.
We teach you to wield it — and when to trust it.
From first connection to your own verified system.
COURSE 01 · FREE — STILL LIFE: STABILITY
Foundations
$0 — forever
How these machines actually work, and how to wire them into your life.
- M0What an LLM actually is — and why it confabulates with total confidence
- M1Setting up Claude properly: web, Projects, Claude Code, your phone
- M2First real workflows — and the first rule of judgment
COURSE 02 · FLAGSHIP — SPACESHIP: MOTION
The Augmented Researcher
$129 founding — one-time, 30-day refund
The full method: every power tool, and the discipline to trust selectively.
- M3Claude Code as your instrument: CLAUDE.md, skills, plugins, hooks
- M4Connect everything: MCP servers, GitHub, subagents
- M5The vocabulary, demystified — and how to judge any tool in six questions
- M6Research that survives scrutiny: quote-and-verify, with real documented catches
- M7Build the system that builds for you — memory, automation, capstone
COURSE 03 · EN ESPAÑOL — GUN: GENERATION
Evaluación rigurosa de IA
Cohorte fundacional en vivo — lista de espera
El primer programa hispano cuyo centro es el juicio, no las herramientas.
- S1Seis sesiones en vivo, en español, sobre tus propios casos
- S2Verificación aplicada: citas, números, fuentes, decisiones
- S3Cupo corto, trabajo real, cero humo
Honest terms, always: no countdowns, no “limited spots” theater, no bonus stacks expiring at midnight. Foundations is free because the basics should be. The founding price is simply the first price. When something isn’t ready, it says so. Each course wears a Life pattern — the still life holds steady, the glider moves, the gun generates. That’s the path.
Built with the method it teaches.
This school is its own first case study.
Every course, every page, every deck here is produced with Claude Code, using the exact workflows we teach — then verified the way we teach you to verify. The animation at the top is a real Conway simulation, credited. Every quote and number is checked against its primary source before it goes up. If it can’t be linked, it isn’t here.
The method has already survived contact with the hardest adversarial domain we know: financial markets. Our sibling research workshop grades every claim by evidence level and publishes its failures by name — tradewiki.org.
FIG. 2 — THE METHOD. THE MODEL PROPOSES; SOURCES DECIDE; YOU RATIFY. THE SECOND LOOP IS WHAT KEEPS THE FIRST ONE HONEST.
- Every factual claim on this page links to a primary source — colophon below.
- The milestones in The Ascent are the ones that held up to a real check — chosen for what’s verifiable, not just what’s viral.
- A model’s confidence is not evidence. We teach you to separate the two — that’s the whole second half of the craft.
- Built in public by one researcher, with the machines he teaches you to wield.
09 — ABOUT
“The most useful thing I’ve learned from eight years of being wrong in public is a method for noticing it sooner. AI just turned that method into a superpower — for anyone willing to learn both halves.”
— ARTHUR ABREUS
Arthur Abreus (@arthurresearch) is a researcher working where epistemology meets the practice of inquiry. He studied finance at the Universidad Complutense de Madrid and spent some eight years in markets before turning to philosophy — a master’s at University College London (UCL), and now further graduate study in epistemology and comparative philosophy. He runs Emergence, an independent markets-research workshop where every claim carries an evidence grade and failed models are published by name.
A confession that doubles as the method: I don’t know your field — and that’s exactly the point. These tools are universal; the same moves carry across science, markets, law, and design. What’s scarce is aiming them at what you know, and checking what they return. So I don’t teach you my expertise — I teach you to amplify yours.
COLOPHON — EVERY CLAIM, SOURCED
- Anthropic Institute (Favaro & Clark), When AI builds itself, 4 June 2026 — “more than 80% of the code we merge…” The paper notes lines-of-code is an imperfect measure and frames the trend as a call for caution.
- Andrej Karpathy, “The hottest new programming language is English”, 24 January 2023.
- OpenAI, Introducing ChatGPT, 30 November 2022.
- TODAY, “A boy saw 17 doctors over 3 years… ChatGPT found the diagnosis” (tethered cord syndrome; confirmed and operated on by a neurosurgeon), September 2023.
- The Nobel Prize in Chemistry 2024 — Hassabis & Jumper for protein-structure prediction (AlphaFold), 9 October 2024; AlphaFold mapped ~200M proteins (DeepMind/EMBL-EBI, 2022).
- Google DeepMind, Gemini officially achieves gold-medal standard at the IMO (35/42, certified by IMO coordinators), 21 July 2025.
- Google DeepMind, AlphaEvolve — improved on Strassen’s 1969 matrix-multiplication algorithm; advanced ~20% of 50+ open problems, 14 May 2025.
- Nature: “AI cracks 80-year-old mathematics challenge — researchers are astonished” (OpenAI model, planar unit-distance / Erdős), May 2026.
- W. Ross Ashby, An Introduction to Cybernetics, 1956, §14/7 (amplifying intelligence); Design for a Brain, 2nd ed. 1960 (ultrastability — “the second loop”).
- J.C.R. Licklider, Man-Computer Symbiosis, IRE Transactions, March 1960.
- Douglas Engelbart, Augmenting Human Intellect: A Conceptual Framework, SRI, October 1962.
- Martin Gardner, “Mathematical Games”, Scientific American 223(4), October 1970 — Conway’s Game of Life (Fig. 1). Glider: R. K. Guy, 1969; Gosper gun: Bill Gosper, 1970.
- Eric S. Raymond, The Glider: A Universal Hacker Emblem, October 2003.
- Frederick P. Brooks, Jr., The Computer Scientist as Toolsmith II, CACM 39(3), March 1996, p. 64 — “IA > AI.”
- Matthias Steup & Ram Neta, “Epistemology”, Stanford Encyclopedia of Philosophy — “the study of knowledge and justified belief.”
- Michael Townsen Hicks, James Humphries & Joe Slater, “ChatGPT is bullshit”, Ethics and Information Technology 26:38, 2024 — LLM output as “indifferent to the truth” (Frankfurtian sense).
- “Epistemic Diversity and Knowledge Collapse in Large Language Models”, arXiv:2510.04226, 2025 — model output less epistemically diverse than a basic web search; RAG mitigates, scale worsens. Preprint; authors per arXiv listing.
- Jason Wei, “Asymmetry of verification and verifier’s law”, July 2025 — many tasks are easier to verify than to solve (with noted exceptions).
- Anthropic, Claude Fable 5 and Claude Mythos 5, 9 June 2026 — generally available; Anthropic-reported SWE-bench Pro 80.3% (vendor-reported benchmark, shown in the launch chart).
- Emergence — markets research workshop: tradewiki.org.
FIG. 1 IS A LIVE CONWAY SIMULATION (B3/S23) — RANDOM SOUP RESOLVING INTO A GOSPER GLIDER GUN. NOT A VIDEO. THE ACTUAL MATHEMATICS.
THIS PAGE WAS BUILT WITH CLAUDE CODE, USING THE METHOD IT TEACHES. CLAIMS VERIFIED 2026-06-09.