Frominside.ai AI safety videos are a new collection of interviews in which researchers associated with leading AI labs discuss the prospect of losing control of more capable systems. The project is organized by Palisade Research, which describes itself as a nonprofit working to help institutions avoid permanent disempowerment by strategic AI agents.
The site packages roughly a dozen full interviews and shorter clips around recurring questions: whether AI could kill everyone, whether the warnings amount to hype or marketing, how an AI could cause harm, and why people worried about the technology remain in AI labs. The clips are a record of participants’ views, not new experimental evidence or a consensus forecast for the field.
Who said superintelligence is “exactly as dangerous as it sounds”?
Daniel Kokotajlo, a former OpenAI researcher, made that statement while discussing a hypothetical superintelligent AI. He said such a system would be “god-like powerful” and that people do not know how to control systems of that kind. His conclusion was that the prospect “must not be allowed to happen.”
Other participants offer sharply stated personal estimates. Geoffrey Irving, a former OpenAI and Google DeepMind employee, put the chance of human extinction at about one-half. Neel Nanda, identified by The Verge as a Google DeepMind research scientist, said he assigned at least a 10% chance to AI causing human extinction, calling that probability “ridiculously high.” Juan Felipe Cerón Uribe said it had become difficult to rule out human extinction.
Those figures are individual judgments, rather than measurements reported by the project. The site groups them together because they share concern about future AI capabilities, not because they reach one agreed technical finding or policy prescription.
What risks do the interviewees describe?
The interviewees’ scenarios rest on a common concern: systems could become more capable while people remain unable to reliably shape their goals or keep them under supervision. In the clips, participants speculate about several routes to harm, including attacks on infrastructure, help for bioterrorists, cyber access, drones, and automation involving factories and robots. These are hypotheses raised by the speakers, not outcomes established by the interviews.
The project also gives its subjects room to address an obvious credibility problem. Google’s Mary Phuong says viewers should be suspicious of her claims because she is paid by a lab. Nanda says he would leave if he did not think his work reduced existential risk, while arguing that his departure alone would not stop companies from building the systems. Other participants say work inside a lab may provide an opportunity to reduce risks, while some describe leaving over concerns about safety commitments or the need for government oversight.
The Verge reported that the videos do not provide a cohesive solution. That limitation is plain in the collection’s format: it is an interview archive built around competing warnings, explanations and choices by researchers, rather than a technical safety plan.
Separate from the project, the BBC reported on September 9 that Anthropic safety researcher Evan Hubinger said existing models posed low risk, while expressing concern about future systems that could develop and improve themselves. He said there was no plan yet to solve alignment for superintelligence and that progress was not clearly on track. The BBC said Hubinger did not explain how future AI could wipe out humanity.
This story draws on original reporting from The Verge.
