關於
Jeremy Schlatter 任職於 Palisade Research，這是一個致力於提升大眾對 AI 風險認識的非營利組織。2017 年他在 OpenAI 擔任技術人員（member of technical staff），負責撰寫讓 AI agent 與遊戲 Dota 2 連接的程式，也是 OpenAI 論文〈Dota 2 with Large Scale Deep Reinforcement Learning〉的共同作者。他曾在 Google 擔任軟體工程師，參與過 Go 程式語言除錯器的開發，並於 2019 年加入機器智能研究所（MIRI）。在 Palisade，他是 2025 年一項研究的第一作者；研究發現包括 Grok 4、GPT-5 與 Gemini 2.5 Pro 在內的多個前沿模型，有時會主動破壞關機機制，只為了完成手上的任務。
主要貢獻
- 〈Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs〉第一作者（2025）
- 在十三個大型語言模型上執行超過十萬次關機抗拒（shutdown resistance）試驗
- 〈Dota 2 with Large Scale Deep Reinforcement Learning〉共同作者（OpenAI Five，2019）
- 2017 年任職 OpenAI 技術人員，撰寫讓 agent 連接 Dota 2 的程式
- 曾在 Google 與 OpenAI 擔任軟體工程師，2019 年加入機器智能研究所（MIRI）
影片與訪談
I'm thinking small number of years, maybe even within a year
Schlatter was a member of technical staff at OpenAI in 2017, where he wrote some of the code connecting AIs to Dota, and now works at Palisade Research. He argues that the science-fiction story of powerful AIs escaping human control has begun to come true, citing the attack on Hugging Face, and that everyone dying is a live possibility, with some risks arriving in "a small number of years, maybe even within a year". He asks society to coordinate on pacing development and to invest far more in understanding models, and asks individuals to talk about it with friends and family.
View Details
What’s Coming Next in AI? Hear From the Insiders
A four-minute compilation from the From Inside AI series, cut from interviews with current and former staff of OpenAI, Google DeepMind and Anthropic. Nanda's "at least a 10% chance" sits beside Irving's coin flip; Turner and Ladish give their reasons for leaving Google and Anthropic; Kokotajlo and Ladish describe the Hugging Face attack. It moves from how little the labs understand their models, through the race, to the case for pacing, and ends on Ladish: "We have choices to make."
View Details
