AI Researchers Speak Plainly on Future Risks in New Video Archive
Palisade Research has published twelve interviews with AI researchers on a new site, frominside.ai.
The series brings together current and former staff from OpenAI, Google and Anthropic. The site offers full videos of each conversation plus short segments grouped by topic, according to coverage by The Verge.
Palisade describes itself as a nonprofit studying what AI systems can do and what drives them. The format is simple. Researchers speak in their own words about how capabilities may develop, how control might be kept or lost, and what risks society could face. Viewers can browse by person or by theme.
Three participants show the range involved. Daniel Kokotajlo is a former OpenAI researcher. Geoffrey Irving worked at OpenAI and Google DeepMind. Neel Nanda is a research scientist at Google DeepMind The Verge.
Kokotajlo is also linked to the AI Futures Project Palisade Research. Jeffrey Ladish interviewed him about the AI 2040 scenario on the Palisade Podcast. That conversation was published on Aug. 25, 2026.
Kokotajlo speaks plainly in the series. He called work on superintelligent AI, systems far beyond human ability, "exactly as dangerous as it sounds" and said it must not be allowed to happen. He also called superintelligent AI "god-like powerful" and said "we do not know how to control it at all" The Verge.
Irving and Nanda put numbers on p(doom), the lab shorthand for the estimated chance of catastrophic outcomes. Irving said the chance of human extinction from AI is "about a coin flip" in his view. Nanda said there is "at least a 10 percent chance AI causes human extinction" The Verge.
Ladish has made similar comments outside the series. He said that "extinction from AI is a very real risk." He also said that "if superintelligent AI agents run society, humans will be at their mercy" Palisade Research.
The broader context here helps place those warnings. Palisade's own test of current models in July 2025 found they were not yet able enough to meaningfully threaten human control Palisade Research. The interviews describe concern about future systems that can plan on their own, remember over time and use software tools, not a claim about systems in use today.
In my view, that distinction shapes how technical readers should use this archive. These interviews are not lab measurements, red-team reports (structured efforts to break a system to find flaws) or detailed safety specifications. They are informed judgments from people close to training and deployment, best read alongside hard data on model behavior, safeguards and monitoring after release.
Looking at what this means in practice, the value lies in comparison. Grouping clips by topic lets engineers, safety researchers and policy staff compare ideas about possible harms, timelines and where safety work could help, without relying on secondhand summaries. For teams writing tests, response plans for accidents, or access rules for powerful models, hearing where uncertainty sits is useful input even when estimates differ widely.
Looking ahead, careful building still offers more than talk of stopping. If highly capable agents ever take on central coordination work in society, control will be settled in code, testing and institutions, not in interviews. A public record of what researchers think now helps define what must be shown to be safe, and that clarity is what makes steady progress possible.
