AI brain rot

Miles Cranmer (PySR author, Cambridge astrophysics professor) wrote a talk for his MPhil students: “How to avoid AI Brain Rot” (3k❤/105k views; full 55-min talk transcribed locally — transcript in cranmer-how-to-avoid-ai-brain-rot-talk). The thesis: AI doesn’t rot your brain directly — it pulls you to the population mean, and the rot is what happens if you let it.

The mechanism: collapse to the mean

• Everyone has AI. Two-three years ago using it was a cheat code; US-adoption studies show even 70-year-olds using chatbots (12% → majority in two years). Standing out is now against the default.
• Slide 10, the talk’s core image: “It pulls you to the population mean” — before AI, work quality is a wide bell curve; with default AI use, everyone collapses into a tall narrow spike (weak side pulled up, strong side costs less effort). The goal of the talk: how to live in the upper tail instead.
• Homogenization is measurable: 25 LLMs × 50 prompts for “a metaphor about time” → essentially two clusters (“time is a river”, “time is a weaver”). Short-story study: AI-assisted stories rated ~8% more novel but dramatically more similar to each other. His assessor’s view: student presentations converged to “the Claude style” — font, colors, everything. Better on average, identical in distribution.
• Who gains: customer-support study — lowest quintile gains most, highest gains least (squashing). Consultant study (750): AI helps on tasks it can do, hurts on tasks just beyond its reach.

The subjective-judgment escape hatch

Where experts still win: tasks needing subjective judgment. Experts+AI 33% vs novices 15% on a software task; novices give up when sessions hit trouble, experts steer the trajectory. His 14-hour autonomous research run (symbolic regression): agent returned “all goals proved, every criterion met” — and the output was “complete garbage” to anyone with taste in the domain (“this is a clean equation, this is a mangled equation” — nonsense). Models are RL-trained on verifiable rewards; subjective tasks have no reward signal, which is also his explanation for why writing got worse: “there’s no equivalent of Lean for writing — unless you simulate a human brain.”

The slogans (keep whole):

• “You are 100% responsible for anything an AI generates — AI can’t go to jail.”
• “Use AI to accelerate yourself, not replace yourself. If you use it to replace yourself, you will be replaced.”
• His delegation test: delegate freely what you can explain and defend without the AI (“nowadays I delegate almost all of my coding — because I can explain it”).

The learning-science core: struggle is the point

• Turkey, 1,000 high-school math students: no-AI 28% on practice, ChatGPT-solutions 42% — but on the exam, no-AI ≈ hints-only, and the solutions group suffered most. The kicker: the solutions group didn’t notice they learned less; the struggle group felt worse and learned more.
• Trio (Python library) quiz: learned-without-AI beat learned-with-AI.
• The granular one: per-student ChatGPT interaction logs — every solution request dropped the test score, every explanation request raised it. The ask is the variable.
• Retrieval-practice crossover: study-4× → 83% at 5 min; study-1× + test-3× → wins a week later. “Desirable difficulties” — struggle is the training stimulus: “if a machine lifts the weight for me, my muscles don’t get stronger. Prompting ChatGPT is like loading the weights onto the machine.”

The practices

1. Socratic loop: “Don’t tell me the answer. Ask me one question at a time. Give a hint only when I’m stuck.” Explain back → AI finds the gap → read → retry. Less fun, better recall, more intuition.
2. Calculator rule: skills you’re not learning + can check → hand over freely; skills you can’t check → learn enough to check; skills you’re supposed to learn → always do yourself first.
3. Single-file visualization demos: “I don’t get why the learning rate matters — build a one-file webpage to visualize it.” Warning: sometimes the demo is just wrong; one tool in the toolbox.
4. Plots over essays (his coding-agent instructions): “You wrote all this code but I don’t have time to read it — make a plot, pretend you’re my student bringing a plot to a meeting, and I’ll ask you about it.” The motion-sensor example: the fit was garbage (“it goes through the floor”) and only the plot showed it. Companion instructions: write the least code that works, always reproduce the bug first, deeply understand what causes it before patching (fix the disease, not the symptom — why he likes ponytail (OSS) and pstack-claude (OSS)). Flowcharts for code architecture (“you click in and go: that’s so dumb, why do they do that”), plots for data analysis.
5. System instructions as your standards: global rules like “when teaching me: quiz me, make me work, don’t hand me solutions” — “it lets you define what good work is.”
6. Ride the wave: “50% comfortable with the newest tool beats 90% comfortable with a two-year-old tool.” The only constant is adaptation; learning-fast is the meta-skill. And the optimism: AI automates pieces → “now let’s solve even harder problems. Raise your ambitions.”

Field answers worth keeping (Q&A)

• “Lights-off software factory” (agents iterating on a codebase from bug reports alone): works on a 3–6-month timeline, then deteriorates — technical debt accumulates until humans must read everything. Current AI lacks long-term architectural vision; human oversight of trajectory remains necessary.
• Software is 3 years ahead of other fields with AI — watch how software evolves to predict your field’s curve.
• On novelty: models are “good at composing ideas — small steps — but going far out of distribution or stepping back for the big picture is harder.” You are still the idea generator; they test and rapidly build.

Kin

• What matters in AI world — the upper-tail doctrine; taste as the thing RL can’t reward (his writing-worse explanation is the taste thesis from the model-training side).
• Vibecoding vs AI engineering — the delegation test (“explain it without the AI”) is the boundary between the two.
• “Don’t be a meat proxy” (gruhn.me, HN thread captured) — the failure mode this talk’s practices prevent.
• Software engineering in AI era — lights-off factory decay; read-all-code-he-writes as PySR’s maintainer mode.