Scott Alexander said he thinks the AI alignment community did not really expect LLMs.

The proposition is that Scott Alexander said this, so it is tested first on attribution and then on substance. True: a transcript or recording from around 2025-04-03 (for example the Dwarkesh Patel podcast with Daniel Kokotajlo) shows Scott Alexander saying, in substance, that he thinks the alignment community did not really expect LLMs. The evidence also has to show the view is reasonable: most prominent alignment writing before about 2019 (MIRI, LessWrong, Bostrom and similar) focused on agentic, reinforcement-learning or other non-language-model routes to AGI and did not centre on large language models. Mostly true: he made the statement, but the words differ in a way that slightly changes the emphasis (for example he said 'some of the community' or 'the timing of LLMs'). Mixed: he made the statement, but the record shows that a substantial, prominent part of the alignment community did explicitly anticipate LLM-like systems as the main path. Mostly false: the statement is attributed to the wrong speaker, or he said something materially different, such as 'expected LLMs but not their capabilities'. False: the transcript shows he did not make the statement, or that he said the opposite.

Set before the research began.

1. E1 AI 2027: Dwarkesh's Podcast with Daniel Kokotajlo and Scott Alexander Zvi Mowshowitz (Don't Worry About the Vase, Substack) · 7 April 2025

   “As Scott says, the alignment community did not expect LLMs, but also we are moving back towards RL-shaped things.”

2. E2 A newcomer’s guide to the technical AI safety field — AI Alignment Forum AI Alignment Forum (post by zeshen) · 4 November 2022

   “Most of the research and publications at the time were from MIRI. They tend to focus on understanding the nature of intelligence and agency i.e. agent foundations, with much of their research work around embedded agency, decision theory, and value learning.”

3. E3 Links For April 2023 Astral Codex Ten (Scott Alexander) · 20 April 2023 Own record

   “I personally am pretty surprised it was just “GPT-4 got released and was very good”.”

E1 is an independent write-up of the Dwarkesh Patel podcast with Daniel Kokotajlo and Scott Alexander. It credits the view to him directly: "As Scott says, the alignment community did not expect LLMs." On substance, E2, an independent guide on the AI Alignment Forum, says that most early alignment research came from MIRI and focused on agency, agent foundations, decision theory and value learning, not language models. That supports the view as reasonable. E3 is Scott Alexander's own post, so it adds only weak context about his surprise at GPT-4. We have no transcript to check his exact words, but E1's wording matches the claim closely.