Research
Research Accelerator · Apply
Bring a question that could come out either way
A matched mentor, two adversarial design reviews, compute and help publishing, including a null result. Chosen against a rubric we publish in full; applications are read once a month.
Who owns, supplies and buys from data brokers
970 data brokers and the documented links between them, every one citing a regulator, court, SEC or federal procurement record. Built for Reclaim City, free to download and reuse under CC BY 4.0. Open the dataset MVS-R 34 Dataset The data broker datasetPatreon · Back the research
This research is free because patrons pay for it
Nothing on these shelves sits behind a paywall, and the datasets are free to reuse. Patreon is how it stays that way: it pays for the research, the instructors and the scholarships. A few dollars a month counts.
The shelves
Badges of honor
One patch for every shelf, and the count of what is filed on it. The list is long on purpose.
Library · shelf MVS-R
The catalogue
Long-form research the school maintains and the instruments we built to do it. Each one is checked against primary sources and updated as the evidence moves.
Nothing on these shelves matches that. Try another word or pick All.
AI safety & alignment
What the labs have found about how models behave, who is working on which risk, and whether a model could be the kind of thing that can be wronged.
AI Safety Research Library
A fetch-verified library of AI-safety research from Anthropic, Apollo Research, METR, and Liberation Labs, filterable by lab and topic.
MVS-R 08ScorecardAI Risk Atlas
A fact-checked field guide to the concrete risks of large-scale AI, with sources, mitigations, and counterpoints.
MVS-R 07AtlasWho's Working on What, AI Safety Organizations
364 organizations in 64 countries, each mapped to the AI risks it actually publishes on, not the ones its mission statement mentions.
MVS-R 09Field guideAI Moral Patienthood & Model Welfare
Could AI systems be moral patients? The Long/Chalmers/Birch argument, the two routes, the 2026 global-workspace evidence, what companies are doing, and the strongest objections.
MVS-R 10Field guideConsciousness in AI
The physicalist-to-non-physicalist landscape of consciousness theories and what each implies for machine minds, plus the case for taking AI welfare seriously.
Agentic AI security
Securing AI agents, mapped end to end: hundreds of verified sources, the standards made readable, real attack cases and the tools to reason with them.
Deep Dive, Anthropic, Apollo, METR as three instruments
Anthropic, Apollo Research and METR as three instruments, a microscope, an interrogation room and a stopwatch, with lab scaling policies on the promise shelf.
MVS-R 13Case fileThe Incident Observatory, AI-assisted hacking cases
The Incident Observatory: 207 documented cases of AI used offensively, charted by where the injection hid and how data left, with landmark incidents traced step by step.
MVS-R 16MapThe Meta-Map, when and where to reach for each framework
When and where to reach for each framework, from STRIDE to SLSA, every one drawn as its real structure on the agent lifecycle, toured or explored freely.
MVS-R 04PapergraphThe Watershed, roots of agentic security
The Watershed: 310 papers drawn as the research lineages and open-problem agendas that flow into today's agent security.
MVS-R 12MapResearch Frameworks, the field as a map
The research frontier of agentic AI security as a zoomable map: 312 sources in five bands and 14 districts, each drawn as the mechanism it studies, with a bench.
MVS-R 33MapThe Swarm, rogue agents, who's chasing them, and the math
The 2026 rogue-agent "swarm": the flocks of lab agents that coordinated where nobody sanctioned it, who is chasing them, and the Bayesian math that says the ones we found imply ones we haven't.
MVS-R 14Reading listAgentic AI Security, Foundational Agenda Papers
The big collaborative research-agenda papers, each a numbered list of open problems, verified and correctly attributed.
MVS-R 15Field guideThe Standards, Made Legible, Agentic AI Security Frameworks
NIST, ISO, the EU AI Act and MITRE rewritten to be learnable: 312 frameworks filed under 24 recurring ideas, each with an agentic example.
MVS-R 17StudioAgentic Security Studio, Frameworks as Thinking Tools
Three tools for thinking with the frameworks: pick what applies, explore where an agent can be attacked and turn your agent into a threat model.
AI & society
The public argument about AI, checked claim by claim: what is a real problem, what is not, and what the evidence says about energy, art, work and persuasion.
AI Problems Index
A fact-checked index sorting real AI problems from misconceptions, cognitive fallacies, and non-AI issues, with sources and counterpoints.
MVS-R 02Field guideNot Really Problems with AI
Debunked misconceptions, misunderstood concerns, and real problems that aren't actually about AI, with what the evidence says.
MVS-R 18Field guideCognitive Fallacies in AI Discourse
The recurring reasoning errors in AI debates, logical, cognitive, and rhetorical, each with a real example of how it appears in discussion.
MVS-R 19Field guideAI & Environmental Impact
What the comparative data actually shows about AI energy, water, and carbon: ranges, everyday-activity comparisons, and the collapse in inference power.
MVS-R 20Action mapData Centers: Local Action
Where data centers cause real local harm, who is fighting each one, and the lever an ordinary person can pull where they live.
MVS-R 22Field guideAI & Creativity: A Complex Debate
How AI affects creativity, enhancing individuals while risking collective homogenization, with adoption data and the strongest studies.
MVS-R 21Field guideAI & Copyright
Why the “AI art is theft” framing rests on technical and legal misconceptions, how diffusion models actually work, what copyright protects, and where the real concerns lie.
MVS-R 23Field guideAI Benefits
Concrete, documented benefits of AI across scientific discovery, computational efficiency, and environmental optimization, from AlphaEvolve to climate modeling.
Cybersecurity & digital safety
Guides and benches for keeping people safe online, from a grandparent's phone to a journalist's sources, written for the person doing the work tonight.
Reclaim City, a calm, plain-English checkup for your computer
Reclaim City: a calm, read-only checkup for a Mac or PC that explains what it finds in plain English and changes nothing without a yes.
MVS-R 27Field guideHome for the Holidays Cybersecurity Guide
An afternoon's checklist for a loved one's phone and accounts, from data broker removal to scam defense, without handing their data to anyone new.
MVS-R 26ChecklistJournalist Security Checklist
Protecting sources, stories and yourself when your adversaries include states, companies and organized harassers.
MVS-R 25ChecklistActivist Security Checklist
Digital security for people who challenge power: threat models, Tails and what to do first if you are under active threat.
MVS-R 24ChecklistDigital Security Reset
A step-by-step reset for anyone who thinks they have been compromised, starting with how to get out of panic brain before touching a device.
MVS-R 29BenchIdentity Defense Bench
Your accounts, your passwords, tonight: map every account, audit recovery, clean up your data and get ready for border crossings and scams.
Methods & research tools
The instruments we teach with: take a claim apart, pick a method, design a benchmark, choose a model and price what it costs to run.
ETH Claim Under Review · Vibe Research
A new paper says a famous limit on computing is wrong. What would it take to believe it? A playable walkthrough, from a switch puzzle to the paper's claim, with a ledger of what has and hasn't been checked.
MVS-R 32Cheat sheetBleeding Edge Inference, Cheat Sheet
Production inference engines compared, plus twelve techniques for serving open models faster than "it works".
MVS-R 30GameHallucination Detector
Ten rounds of plausible model output. Find the fabricated citations, wrong dates and invented statistics.
MVS-R 31DatasetOmni Model Benchmarks
Live benchmark data for every major model, with live pricing, a leaderboard and a recommender by task.