开启时会模糊预览图，关闭后正常显示
I wrote a lesswrong post attempting to explain the core part of the hard problem of consciousness and what confuses me about it. Interested for people to read and tell me if they think I'm right. See below
https://t.co/htwwJNORzf
Why, when and how ML models generalize is not just the deepest question in ML, but one of deepest questions humans have ever encountered in any sphere of thought.
Interests: AI (Safety), meditation, philosophy, mathematics, algorithms If I say something you disagree with, please dm or quote tweet. I love to argue!
Opus 4.5 is noticeably better at conversation and intent. I always ask models "Tell me if anything I said is unclear or you need more information", and they never follow up on it (and then proceed to do some dumb stuff), but Opus 4.5 will follow up when its required.
Interests: AI (Safety), meditation, philosophy, mathematics, algorithms If I say something you disagree with, please dm or quote tweet. I love to argue!
PSA that current AI sycophancy is quite heavy handed but will get better and subtler
Interests: AI (Safety), meditation, philosophy, mathematics, algorithms If I say something you disagree with, please dm or quote tweet. I love to argue!
If token space was an information bottleneck for reasoning LLMs, you'd expect RL trained models to learn to use all tokens with high frequency. But when I read cots, they seem narrow, utilizing a smaller space of tokens than base models, or the model output.
Interests: AI (Safety), meditation, philosophy, mathematics, algorithms If I say something you disagree with, please dm or quote tweet. I love to argue!
RT @BjarturTomas: Ron's face pulled into a scowl. "If you don't like Quidditch, you don't have to make fun of it!" "If you can't criticise…
Interests: AI (Safety), meditation, philosophy, mathematics, algorithms If I say something you disagree with, please dm or quote tweet. I love to argue!