關於
Mary Phuong 是 Google DeepMind Frontier Safety「失控風險」（Loss of Control）團隊的研究科學家，2022 年加入至今。她目前的工作是 AI 控制（AI control）：確保那些可能未對齊、在內部部署的模型，即使有意為之，也無法造成嚴重傷害或暗中破壞。她先前主導危險能力評估，測試 Gemini 模型在說服、資訊安全、自我擴散與自我推理方面的能力，之後也評估隱匿能力與情境覺察。她在奧地利科學技術研究所（ISTA）取得博士學位，指導教授為 Christoph Lampert，2021 年的論文探討深度學習中的「欠規格」（underspecification）。她也在 MATS 計畫中指導研究者。
主要貢獻
- 〈Evaluating Frontier Models for Dangerous Capabilities〉（2024）第一作者，在 Gemini 1.0 上試行一套全新的危險能力評估
- 〈Evaluating Frontier Models for Stealth and Situational Awareness〉（2025）第一作者
- 〈GDM AI Control Roadmap〉（2026）第一作者，這是一份針對可能未對齊之 AI 代理人的內部安全藍圖
- 共同撰寫〈Model Evaluation for Extreme Risks〉（2023）與〈Goal Misgeneralization〉（2022）
- 與 Marcus Hutter 合著〈Formal Algorithms for Transformers〉（2022），以精確的數學語言整理 Transformer 架構
影片與訪談
Google DeepMind Safety Researcher: “Be Suspicious of What I'm Saying”
Phuong, on Google DeepMind's safety team since 2022, works on loss-of-control risk and describes AI as an alien mind whose motivations we have few guarantees about. She reports colleagues saying models make them far more productive but also leave them feeling redundant, and explains how models that realise they are being tested can game alignment evaluations. Asked why she stays, she tells viewers they "absolutely should be suspicious" of her because the lab pays her, and says she would quit if she could no longer help from inside. She calls how well we will solve alignment "basically a coin toss".
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
