關於
Juan Felipe Cerón Uribe 從 2022 年起在 OpenAI 擔任研究工程師。他在哥倫比亞波哥大長大,就讀安地斯大學(Universidad de los Andes),後來為了加入 OpenAI 移居美國。他一開始參與的工作後來成為 ChatGPT。現在他是 OpenAI Safety Systems 團隊的研究者,研究對抗式穩健性訓練與自動化紅隊測試,其中包括防止 ChatGPT 協助他人製造生物武器。他曾以個人身分公開談到前沿實驗室承受的競爭壓力:每一代新模型往往在風險還沒弄清楚之前就推出。
主要貢獻
- 2022 年加入 OpenAI 後參與後來成為 ChatGPT 的專案,也列名於 GPT-4 Technical Report 的貢獻者(2023)
- 與 Nat McAleese、Jan Leike 等人共同撰寫〈LLM Critics Help Catch LLM Bugs〉(2024),訓練模型批改程式碼
- 共同撰寫〈IH-Challenge〉(2026),一個用來改善前沿 LLM 指令階層(instruction hierarchy)的訓練資料集
- 共同撰寫〈GPT-Red: Automated Red Teaming via Self-Play at Scale〉(2026)
- 負責防止 ChatGPT 協助製造生物武器的防護措施,並擔任 MATS OpenAI 安全方向的導師
影片與訪談
OpenAI Engineer: AI Labs Are “Racing Each Other Kind of Blindfolded”
Cerón Uribe, a research engineer at OpenAI since 2022 who now works on stopping ChatGPT from helping people make bioweapons, says frontier labs are "racing each other kind of blindfolded". He describes pressure to ship flowing down from leadership, and says talk of existential risk at OpenAI has given way to product and revenue as the company grew, with what he calls a loss of academic freedom. He argues that only internationally coordinated pacing counts, and that pacing "gives you time to do homework, but it doesn't do your homework". He stays because his work still reduces large-scale risk, but cannot say that will hold next year.
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思想連結
Leopold Aschenbrenner
思想對照Situational Awareness LP 創辦人兼投資長
兩人面對的是同一個質疑——如果美國的實驗室放慢腳步,中國就會勝出——卻給出相反的答案。Aschenbrenner 的答案是「領先」:自由世界必須贏下這場競賽,而足夠的領先幅度可以「兌現」一部分,換取把安全做對的時間;國際安全條約在他看來不切實際。Cerón Uribe 則說「沒有單方面的放慢」,唯一算數的是「國際協調下的放慢」。兩人真正的分歧在於,安全所需的時間從哪裡來:是一方先存下的領先,還是雙方約好一起放慢。
situational-awareness.ai · youtube.com
Daniel Kokotajlo
思想對照AI Futures Project 執行主任
兩位 OpenAI 內部人,診斷相近,選擇相反。Kokotajlo 說,他是在不再相信公司會在 AGI 到來時負責任地行事之後離開的。Cerón Uribe 被問到為什麼不乾脆辭職,回答說 OpenAI 仍讓他做得到能「實質降低」大規模風險的工作;同時他也坦言,隨著工作被拉向產品,研究的自由少了,而他無法保證明年的答案還是一樣。放在一起讀,這就是身在實驗室裡的人僅有的兩種回答:離開,或留下來做事。
greaterwrong.com · youtube.com