iamtrask reframes AI safety debate around unilateral vs collective control

On September 14, OpenMined founder and former OpenAI researcher Andrew Trask (iamtrask) reframed the AI safety debate in a tweetstorm: the open vs closed source dispute is not the real issue—unilateral control of AI is. His conclusion is that collective control is the only way forward. The next day, aronchick built on this with a more concrete governance proposal.

Confirmed

iamtrask argued that 99.999% (written as 99.9% in some posts) of the AI safety debate circles around two questions: who should unilaterally control superintelligent AI, and how to ensure AI does not control itself.
He identified three forms of unilateral control: closed AI controlled by a single company; open AI controlled by anyone holding a copy; and runaway AI controlled by itself.
He claimed no one can be trusted to unilaterally control superintelligence, and the only answer is 'collective control'—the secret mechanism behind democracy and free markets.
As a technical argument, he said collectively controlled models (model fusion, ensembles, routers) already beat frontier single models on accuracy and cost, but rarely appear on leaderboards due to attention bias.
He added that collectively controlled AI is easier to regulate with far less regulatory capture risk, and promised a longer post to follow.
On September 15, aronchick, citing Ramez's point that intelligence should not be controlled by the few, proposed that no single entity—company, government, or coalition—should control intelligence, and that all organizations training frontier models (including open-source ones) should undergo third-party audits, with passing auditors earning a 'dolphin safe'-style label.

Why it matters

The framing turns the long-running open vs closed debate into a question of power structures, offering a new analytical lens.
The claim that collective control is more accurate and cheaper grounds the position in economic and performance incentives, not just values.
The easier-to-regulate argument and aronchick's third-party audit and safety label proposal are directly relevant to policy discussions on superintelligence governance.

2026-09-14 ~ 2026-09-15 · 8 related posts

Primary sources

iamtrask: The Real AI Safety Problem Is Unilateral Control, Not Open vs Closed — iamtrask · 2026-09-14 07:44 (UTC+08:00)
iamtrask: Collectively Controlled AI Is Easier to Regulate With Less Regulatory Capture — iamtrask · 2026-09-14 07:46 (UTC+08:00)
A proposal for decentralized AI: third-party inspection and a 'dolphin safe' label — aronchick · 2026-09-15 02:43 (UTC+08:00)

iamtrask: Open vs closed source isn't the issue — unilateral control of superintelligence is — iamtrask · 2026-09-14
[source] iamtrask: The Real AI Safety Problem Is Unilateral Control, Not Open vs Closed — iamtrask · 2026-09-14
iamtrask Maps Three Forms of Unilateral AI Control: Company, Copy Holder, the AI Itself — iamtrask · 2026-09-14
OpenMined's Trask: AI safety debate boils down to who unilaterally controls superintelligence — iamtrask · 2026-09-14
iamtrask: Nobody Can Be Trusted With Unilateral Superintelligence Control — Only Collective Control Works — iamtrask · 2026-09-14
iamtrask: Model Fusion and Ensembles Beat Frontier AI on Accuracy and Cost — iamtrask · 2026-09-14
[source] iamtrask: Collectively Controlled AI Is Easier to Regulate With Less Regulatory Capture — iamtrask · 2026-09-14
[source] A proposal for decentralized AI: third-party inspection and a 'dolphin safe' label — aronchick · 2026-09-15