Helen Toner

AI models exhibit extreme unevenness in capabilities—excelling at complex reasoning tasks like PhD-level questions while simultaneously failing at simple visual tasks like counting line intersections—and this jaggedness pattern will likely persist.

“Author of "Taking Jaggedness Seriously" in Substack”

“Noted that timelines to superintelligence are uncomfortably short, defining 'minimum viable superintelligence' and speculating arrival within 5-10 years.”

“Then-board-member of OpenAI who faced criticism from CEO Sam Altman for coauthoring an AI safety paper that implicitly criticized the company.”

“Suggested using Hugging Face incident as topic for Trump-Xi discussions and as evidence of importance of US-China AI safety collaboration.”

“Board member who received Sutskever's memos confirming her concerns about Altman's trustworthiness and voted to fire him, later losing her seat upon his reinstatement.”

“Co-author of AI research study examining whether agents can conduct open-ended research.”

“Pointed out insiders have expected incidents like this for a long time and that internal models are a major threat source.”

“Discussed in a 2024 podcast how OpenAI executives shared evidence with the board of Sam Altman lying and being manipulative in various situations.”

“Wrote essay about how long AI timelines have shortened due to recent progress.”