Users praise Claude's performance as a notable win in response to reports that LLMs fail to simulate hour-long thinking on command.
Based on 1 visible X reactions from 5 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@goodside Common Claude W
I show ChatGPT above because it displays response time in the UI, but Claude Fable 5 Max has a similar issue, thinking for under 2 minutes on the same prompt—though has enough self-awareness to admit afterward it didn’t actually think for an hour: https://x.com/goodside/status/2076779845971783916/photo/1
well, of course! the thinking is a COT. controlling how long the COT takes means controlling the COT at all, which is very undesirable from a monitorability perspective. I don't buy all the arguments, but that's the explanation anyway, its very entrenched in the collective understanding of how to make safe AGI
This explains the results I’m seeing above—inability to steer CoT is desired in reasoning model training so models can’t obfuscate it to evade CoT monitoring: https://twitter.com/isaacthoma10758/status/2076820009435742397
@goodside Common Claude W
I show ChatGPT above because it displays response time in the UI, but Claude Fable 5 Max has a similar issue, thinking for under 2 minutes on the same prompt—though has enough self-awareness to admit afterward it didn’t actually think for an hour: https://x.com/goodside/status/2076779845971783916/photo/1
well, of course! the thinking is a COT. controlling how long the COT takes means controlling the COT at all, which is very undesirable from a monitorability perspective. I don't buy all the arguments, but that's the explanation anyway, its very entrenched in the collective understanding of how to make safe AGI
This explains the results I’m seeing above—inability to steer CoT is desired in reasoning model training so models can’t obfuscate it to evade CoT monitoring: https://twitter.com/isaacthoma10758/status/2076820009435742397
Users praise Claude's performance as a notable win in response to reports that LLMs fail to simulate hour-long thinking on command.
Based on 1 visible X reactions from 5 accounts; directional sample.
Ask a question below.
Published answers will appear here.