Roon

Pseudonymous X poster (@tszzl), widely known to work at OpenAI and one of the most-read voices from inside a frontier lab. The account itself names no employer. Its bio reads “ceterum censeo we must pace the frontier of global machine intelligence progress”.

2026: Alignment as Plan A

During the week of the Against SURA debate, roon set out why staff at frontier labs treat technical alignment, not an international pause, as the main plan. The posts were read through the fxtwitter API on 2026-11-05, because X refuses automated fetches.

On 2026-11-02 roon posted that “Plan A has always been to solve alignment”, by the same incremental methods as the rest of modern AI rather than by first enhancing humans. The next day Will Fithian, a UC Berkeley statistician, quoted it: “Making ASI safe by ‘solving alignment’ is a delusion”, because it gives no plan “to stop the world’s criminals, terrorists, and tyrants from aligning machines to their own”. This is the Pivotal act objection: aligning your own AI does not stop anyone else’s.

Roon’s reply, 2026-11-04:

to be clear my ideal case is that countries think of this as more dangerous than nuclear weapons and create international agreements around it

I consider solving sufficient alignment plan A because I don’t expect global mutually assured coordination to happen in short order, without major demonstrations of destructive force, and long periods of dangerous weapons accumulation and brinksmanship, which may or may not come before it’s too late

there is no plan to stop everyone from having this; it does not seem possible

every bad thing will happen, there is no point trying to reason about a world where everything goes according to plan:

• someone will hook up ai to the internet
• hacker groups will use AI to cause damage on the internet
• a non-alignment pilled group somewhere will develop intelligence they cannot control and lose control of it very soon
• a rogue AI will improve itself on some dimensions

we want to create a world robust to all this by investing in defense earlyD [sic] depending on the offense/defense balance of future technologies we can hopefully coexist with criminals and terrorists and rogue AIs having some minor amounts of compute

we better solve alignment

A treaty would be best but will come too late, nobody can be kept from having powerful AI, and every failure should be assumed. So the hope is early defense and a favorable offense/defense balance. The reply does not say how solving alignment answers Fithian’s question, which depends on How hard is alignment.

Roon’s “plan A” is close to Scott Alexander’s Plan C. In the AI Futures Project scheme that Scott uses, Plan A is the global pause; see Plans for the AI transition.

Related: Can AI progress be paused, Should safety-focused labs keep building.