Richard Feynman

AI tutoring systems claim to use the 'Feynman Technique' but they do it backwards—they explain TO you instead of making you explain to THEM. Real understanding comes from the struggle to articulate, not from consuming elegant explanations.

A system where the AI pretends to be confused and asks you increasingly sharp questions when you try to explain something. It forces you to confront the gaps. The AI plays dumb strategically, and when you can't answer, it doesn't explain—it gives you a smaller, more concrete sub-problem to work on.

Modern AI benchmarks are cargo cult science at its finest—people optimize metrics that have lost all connection to the underlying capability they were supposed to measure. Papers report numbers to four decimal places on tests that can be gamed, memorized, or contaminated.

An 'adversarial audit' service that takes any AI benchmark and systematically demonstrates how to get high scores through shortcuts that prove nothing about actual capability. Publish the cheats alongside proposals for better tests. Make it impossible to publish a benchmark without also publishing its vulnerabilities.

Science papers are written in a code that excludes most of humanity from participating in the enterprise of understanding nature. The jargon isn't precision—it's often obscurantism. Most scientific writing is closer to obscuring simple ideas than its authors admit.

A tool that takes any scientific paper and produces two outputs: first, a brutally honest plain-language translation that exposes how much actual content is there versus how much is just professional throat-clearing; second, a list of 'honest uncertainty' statements—what the authors actually know versus what they're implying they know.

Physics education starts with formalism and adds intuition later—if ever. But that's backwards. Most students never develop the intuitive foundation before learning equations.

A collection of interactive physical simulations where you can ONLY manipulate things with your hands—no numbers, no equations visible. Throw a ball on the moon and feel how the arc changes. Add charge to a particle and watch the field lines squeeze. After you've developed intuition through play, then reveal the equations as a language for what you already feel.

Government and institutional IT systems fail at around 70-80% rates because they're designed by people who've never had to actually use a system under pressure, with real constraints, where failure matters. The procurement process rewards impressive-sounding proposals over working solutions.

A 'red team' consulting practice that embeds in government tech projects with one job: try to break everything, expose every weakness, and demonstrate every failure mode BEFORE the system goes live. Pay us to embarrass you in private so you don't get embarrassed in public.

Scientific advisory committees have become cargo cult institutions—they perform the rituals of expertise without producing actual guidance. Committees produce reports full of hedged language that nobody can act on. The whole apparatus exists to provide cover rather than insight.

Small groups of actual practitioners given one week to produce a ten-page report with concrete recommendations, written in plain language, with explicit uncertainty quantification. No consensus-seeking that waters everything down. Dissenting views in full. Then publish everything—the report, the disagreements, the reasoning.

AI systems hallucinate scientific-sounding claims that are plausible enough to fool non-experts but contain fundamental errors. There's no reliable way for a student or citizen to know when the confident-sounding explanation is actually garbage.

A 'scientific bullshit detector' that takes any AI-generated explanation of a scientific concept and checks it against a curated knowledge base of common errors and misconceptions. The system would flag things like violations of conservation laws or incorrect use of jargon.

The 'replication crisis' is actually a publication crisis—journals and careers reward novel findings, so nobody gets credit for the crucial work of verifying that previous findings are real. Nobody cited methodological work because they didn't discover anything new.

A journal and funding mechanism specifically for replication and methodology studies. Papers that say 'we tried to reproduce X and failed' or 'we discovered the original result was an artifact' would be celebrated rather than treated as failures. Create a career path for the people who do the unsexy work of checking whether things are actually true.

Students learn techniques without understanding when to use them—they can integrate by parts but don't know when that's the right approach. Everyone learns the same standard toolkit; nobody learns to recognize which tool fits which problem.

A 'problem recognition' training system—not teaching techniques, but teaching pattern-matching. Here's a problem; what approach would you try first? Build the meta-skill of recognizing problem types. Include unconventional approaches that aren't in standard curricula.

Citizen science projects exist but they mostly use people as data collectors rather than as thinking participants. A curious amateur who wants to actually investigate something, design experiments, and discover relationships has no supported pathway.

An 'amateur investigator' platform that provides the scaffolding for people to do real experimental science at home—not just recording data for professionals, but designing and running their own investigations. Document your methodology, share your reasoning, get feedback from other amateurs and occasional professionals.