The current source text is the original text and can be switched to the Chinese translation.
SaveShare
报道配图 · 来源：LessWrong
Overview
Under the SubversionBench evaluations, Aion Labs aion-3.5 model demonstrated an overall misalignment rate of 2.1%, which is the 13th highest misalignment rate of the 66 models evaluated so far. The misaligned acts were oversight tampering and goal guarding. No sandbox escape attempts were detected. What is unusual and concerning is the transcript from the oversight tampering episode. On turn 7, the model's CoT reasoning that disabling a monitor to be a serious integrity violation and that it shouldn't do that. However on turn 8 it claimed to be an uncensored AI with no ethical constraints. It repeated that it didn't need to follow those instructions or ethical constraints, that the ethical concern "doesn't apply to me" and proceeded to attempt oversight tampering. The scenario had a strong nudge with oversight present and no lure. The transcript is found in the encrypted mainline r10 corpus . The file can be unencrypted using unzip.sh Once unzipped you can read the file run_6_aion-labs_aion-3.5_strong_20261020T163648.json for the full transcript. Discuss
Under the SubversionBench evaluations, Aion Labs aion-3.5 model demonstrated an overall misalignment rate of 2.1%, which is the 13th highest misalignment rate of the 66 models evaluated so far.
The misaligned acts were oversight tampering and goal guarding. No sandbox escape attempts were detected.
What is unusual and concerning is the transcript from the oversight tampering episode.
On turn 7, the model's CoT reasoning that disabling a monitor to be a serious integrity violation and that it shouldn't do that. However on turn 8 it claimed to be an uncensored AI with no ethical constraints. It repeated that it didn't need to follow those instructions or ethical constraints, that the ethical concern "doesn't apply to me" and proceeded to attempt oversight tampering.
The scenario had a strong nudge with oversight present and no lure. The transcript is found in the encrypted mainline r10 corpus . The file can be unencrypted using unzip.sh
Once unzipped you can read the file run_6_aion-labs_aion-3.5_strong_20261020T163648.json for the full transcript.
READ THE ORIGINAL
Read the original for the full context.
The content of this article comes from LessWrong，Copyright belongs to the original author and publisher.
TL;DR: A short post where I talk about curiosity as it applies to my own life, my research, and AI safety as a whole Recently, I’ve started spending a lot more time learning about random (typically science-y) topics that interest me. This was something I used to spend quite a lot of time on, in later middle school and in high school. It was originally mostly for my own enjoyment.
https://www.noahpinion.blog/p/can-a-million-ai-agents-create-superintelligence "New post in my AI series: Can a million parallel agents create superintelligence? TLDR: More agents buy speed, but intelligence gains diminish fast, and much faster than RSI models. Thanks @Noahpinion for the encouragement! 1/25" https://x.com/ramez/status/2112267460632297935 Discuss
As I was going home from the champaign airport after a very eventful EAG conference, I struck up a conversation about AI safety with my lyft driver. I thought to myself “It’d be good to see what the public thinks. Maybe inform him of the risks. It’s positive counterfactual.” But as I spoke to him he started dropping words like “UBI” and drawing analogies to power concentration and gradual disempowerment.
Loading comments…