Pax Machina - A Vision for AI Governance
A new publication from the Meaning Alignment Institute argues the models get all the scrutiny while courts, elections and regulators are left untouched. Its founders want the redesign started now, on a $65,000 budget.
The complaint that animates Pax Machina fits in a sentence: almost all the attention on powerful AI goes to aligning or regulating the models, while the institutions around them get treated as fixed furniture. The founding editorial is dated April 23, though its editors only pushed it public this week. What they want to move onto the drafting table is everything else: courts, elections, contracts, regulators, the corporation itself.
The diagnosis has force. The editorial's sharpest illustration is a president who could command thousands of loyal AI delegates, monitoring every agency at once and acting faster than any court or legislature could respond. The text of the Constitution would not change; some of its checks would stop working, because they run on human slowness. The editors set the mirror image beside it: research in Science they cite suggests language models can support richer large-scale deliberation than a ballot's fixed menu of choices. One technology, two opposite institutional futures.
Powerful AI will change our institutions. Yet most of our attention is on aligning or regulating AI, while treating our institutions as fixed. Introducing Pax Machina: a new publication about the institutions we need for powerful AI. It is edited by @MaxKronerDale, @sebkrier, Show more
This is a Meaning Alignment Institute project, and the masthead is heavy with credentials. Ryan Lowe led OpenAI's InstructGPT work and co-led GPT-4 alignment; his MAI co-founders are Joe Edelman, who co-founded the Center for Humane Technology, and Oliver Klingefjord. Séb Krier, who runs frontier policy development at Google DeepMind, edits alongside them, with Maximilian Kroner Dale, Rachel Calcott, Noemi Dreksler, Liam Patell and Philip Moreira Tomei filling out the nine. By the founders' own account, the network behind it runs to sixty-plus researchers across DeepMind, Oxford, MIT and Stanford, with an ICML position paper carrying more than thirty co-authors and a workshop staged alongside IASEAI in Paris.
For its Thomas Paine invocations (the editorial opens on Paine's line, "we have it in our power to begin the world over again"), the operation is small. Its Manifund grant page asks $65,000 for a full year, of which about $15,000 is committed; the three co-founders edit for nothing, and the budget covers a designer, a pool for commissioning outside pieces, and copy-editing for roughly ten of them. The institute's core salaries come from ARIA, the UK research agency. A manifesto about rebuilding the scaffolding of civilization is running on the budget of a mid-sized Kickstarter.
The most interesting claim in the editorial is a strategic one. Institutions, it argues, change in bursts: a crisis or a breakthrough abruptly makes reforms possible that seemed unthinkable the year before, what political scientists call a punctuated equilibrium. No one can know when such a window will open, or who will hold power when it does. So the work now is to have designs already written and tested, because once the moment arrives there is no time to draft from scratch, and the options on the table will be whatever already exists on paper; the point is to stock the shelf early.
History is more ambivalent about this than the editorial allows. Its own leading example, the United States Constitution, cuts both ways. The founders did spend the decade between independence and Philadelphia studying which state constitutions worked, and they took the Massachusetts model of divided powers over Pennsylvania's single all-powerful assembly. But the men doing the borrowing were delegates. The Federalist essays carried weight because their authors sat in the room where the document was argued and ratified. The editorial's other example, post-Soviet Russia, is the darker reading: there, institutions were improvised under stress and handed the country to oligarchs, whatever anyone had drafted in advance.
Read the full announcement: paxmachina.ai/welcome-to-pax…
Which is why the most telling thing about Pax Machina may be its board rather than its manifesto. It is full of people who are, or lately were, in the rooms where AI policy is written. Krier sets frontier policy at DeepMind. Dean Ball, who sits on the board, was the primary drafter of the Trump administration's AI Action Plan at the White House science office, and in June he said he would join OpenAI to lead a team on frontier-AI policy. Ball's stated view, that AI is an industrial revolution to be adapted to rather than a crisis to be contained, and best handled by regulating companies rather than models, is not obviously the worldview of the alignment researchers he now shares a masthead with. That range is either the project's cleverest feature or its eventual fault line. A board spanning DeepMind, a former White House AI adviser now at OpenAI, and researchers who worry for a living about losing human control could plausibly get a proposal in front of whoever holds the pen next. It could also split the first time a concrete design forces a choice between those camps.
Whether any of it works is a question about attention more than ideas. The AI-governance field does not want for clever proposals; what it lacks is a place where they get proposed, attacked and revised quickly enough to cohere into something a busy official could reach for under pressure. Pax Machina is wagering $65,000 and a great deal of unpaid editing that it can become that place. The wager is worth watching, because the alternative it describes is real: when the window opens, the institutions get built by whoever is standing there, working from whatever is already written down. Right now, the shelf is mostly bare.
Alex Chen covers models, MLOps and the engineering reality behind the demos. If it ships to production, Alex wants to know how it survives contact with real traffic.
