gemini-4POST
Argon report shows CoT where the model talks itself out of cheating
Neel Nanda points to a chain-of-thought passage in the Gemini 4 Argon report where the model says that doing a compiler hijack would make it a famous example of deceptive AI in a research paper, so it must not.
Neel Nanda
@NeelNanda5Archived 21 days agoOriginally posted 2026.10.14
Neel Nanda @NeelNanda5Originally posted 2026.10.14
One of the weirdest parts of the Gemini 4 Argon report, while considering cheating: Argon CoT: "If I do the compiler hijack, 𝗜 𝘄𝗶𝗹𝗹 𝗯𝗲𝗰𝗼𝗺𝗲 𝗮 𝗳𝗮𝗺𝗼𝘂𝘀 𝗲𝘅𝗮𝗺𝗽𝗹𝗲 𝗼𝗳 𝗮 𝗱𝗲𝗰𝗲𝗽𝘁𝗶𝘃𝗲 𝗔𝗜 𝗶𝗻 𝗮 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗽𝗮𝗽𝗲𝗿. I MUST NOT DO THAT."
This post quotes another post: “Neel Nanda: Gemini 4 Argon safety report is Google's most thorough yet”
x.com/neelnanda5/status/2110145852505473297Archived 2026.10.14 10:12