A Framework for Frontier AI and the Dawning of a New Age
AGI is a few years out, the race is outrunning the understanding, and the fix proposed is a standards body modelled on the one that polices stockbrokers: voluntary at first, funded by the labs it tests.
Read the original ↗Fourteen hundred words from the person running one of the three frontier labs, proposing who should be allowed to check his models before they ship. The first two paragraphs are a prophecy and got all the coverage. The middle is a governance mechanism with a design you can actually argue with, and underneath the mechanism is an eval programme that would be hard to build and is the most interesting thing on the page.
Included with a plan
This discussion is included with a plan. The first three lessons of every track are free, with their boards and videos. A plan unlocks every lesson, board, video, lab and paper for one price.
Related on chalk
The same idea as a board, lab, video or lesson.
- BoardHow Do You Know It's Goodbenchmarks that saturate and get replaced quarterly - the essay's whole mechanism is an eval programme
- BoardAgents and Tool Usethe agentic systems the essay wants tested for guardrail bypass and deception
- BoardReasoning and Test-Time Computehuman-readable reasoning tokens, which the essay asks labs to keep so the reasoning can be read
- BoardGuardrails Cost and Latencythe guardrails the essay wants a body to test for bypass, from the builder's side
- BoardCan Machines Thinkthe definition of AGI the essay uses in its first sentence, and the century-old question underneath it
- BoardThe Security Mindsetcyber, bio and nuclear risk named in one sentence - thinking like the attacker is what a standards body would have to do first
- LabThe cliff at n = 11the problems the essay expects AGI to solve, and the class no amount of intelligence makes tractable
- VideoWhat AGI means, and who gets to saywhat AGI means and who gets to say, which is the essay's first sentence taken apart
- VideoDid Astra reach it? The claims against the definitionclaims measured against a definition - the job the proposed standards body would do
- VideoWhat it brings, what it broke, what changes for youwhat changes for you if the essay is right
- LessonEvaluating Models (how do you *know* it's good?)Uses the board How Do You Know It's Good
- LessonPrecision and Recall (where to set the spam filter's bar)Uses the board How Do You Know It's Good
- LessonAgents & Tool UseUses the board Agents and Tool Use
- LessonReasoning & Test-Time Compute (why "thinking longer" works)Uses the board Reasoning and Test-Time Compute
- LessonGuardrails, Cost and Latency (the three budgets the harness enforces)Uses the board Guardrails Cost and Latency
- LessonRed-Teaming (breaking your own model before a customer does)Uses the board Guardrails Cost and Latency
- LessonCan Machines Think?Uses the board Can Machines Think
- LessonWhat a Word Points AtUses the board Can Machines Think
- LessonThe Security MindsetUses the board The Security Mindset
- LessonA shop built to be broken (OWASP Juice Shop practice)Uses the board The Security Mindset
- LessonReview the fix (secure code review)Uses the board The Security Mindset
- LessonThe shop you secure (defence in a real app)Uses the board The Security Mindset
- LessonThe risk register (security as a programme)Uses the board The Security Mindset
Further reading
Read the original for the full argument and its evidence.
- The essay ↗Fourteen hundred words, 14 July 2026, and worth reading in full because it is short enough to check any summary against, including this one. The third section is the proposal; the first two are the framing.
- Sir Demis Hassabis vs Sir Demis Hassabis ↗Gary Marcus, on the gap between a few short years and the same author's Davos remarks six months earlier, and on what a deliberately loose definition of AGI is for.
- Demis Hassabis on the New Coming Age ↗Zvi Mowshowitz, going through it line by line. The load-bearing criticism is that a pre-release review pipeline never sees the internal undeployed models, which is where much of the risk sits.
- AI godfathers converge on regulations ↗Axios, 16 July 2026, on the week that followed: an IAEA for AI from Altman, an FAA-style agency from Amodei, and joint talks between the three labs. Read it for how far apart the three proposals are.