Brundage and Delip Rao Clash Over AI Misuse Predictions and the Merits of Scaling
Former OpenAI policy lead Miles Brundage and researcher Delip Rao clashed over whether "today's AI could have been predicted back then," with the debate centering on the predictive value of a 2018 report and the historical merits of scaling laws.
Confirmed
- Brundage recalled that when he and his co-authors published "The Malicious Use of Artificial Intelligence," scenarios like automated hacking, spear phishing, and lethal autonomous drones were mocked as "science fiction"—and have now become reality.
- Delip Rao fired back in a series of posts, calling Brundage's retrospective claims of successful prediction "intellectually dishonest," arguing from technical history: transformers only arrived in 2017, 2018-era decoder LLMs did not scale automatically, instruction alignment took shape around 2020, and key capabilities like function calling also emerged around 2020. Anyone claiming they could predict today's AI with 2018-era knowledge is making things up.
- Rao went further, arguing that scaling was necessary but not sufficient, that OpenAI's faith in scaling laws was misplaced, and that its success relied on methods invented later (mostly outside OpenAI) that happened to be compatible with "blind scaling"—a matter partly of luck.
- Brundage responded with clarification: the 2018 report discussed possibilities judged "credible within 5 years" based on proof-of-concept and trend analysis, not long-term point predictions; he doesn't think every judgment back then was accurate, and stressed that what he opposes is dismissing misuse risks.
Why it matters
- The debate reflects a deep divide in the AI community over "crediting the scaling thesis": whether scale drove today's AI capabilities, or whether methodological breakthroughs like instruction alignment and function calling were the key—directly shaping how we frame predictions about the future and where to invest resources. - For policy research, "which misuse predictions actually came true, and how" affects the credibility and prioritization of AI safety and governance agendas.
2026-09-07 ~ 2026-09-07 · 5 related posts
Primary sources
- [source] Brundage revisits 2018 malicious AI report: the sci-fi scenarios happened — Miles_Brundage · 2026-09-07
- Delip Rao: in 2018, claiming you foresaw today's AI was sci-fi imagination — deliprao · 2026-09-07
- [source] Miles Brundage: my 2018 report discussed 5-year plausibility, not predictions — Miles_Brundage · 2026-09-07
- Delip Rao accuses ex-OpenAI's Miles Brundage of dishonest victory-lapping — deliprao · 2026-09-07
- [source] Delip Rao: scaling was necessary but not sufficient, OpenAI got lucky — deliprao · 2026-09-07