frotaur

10 ms·

1. 121.▲
   by frotaur 2y ago

   Tried reading the 'low-dimensional diffusion' one. Not an expert on diffusion by any means, but the very premise of the paper seems like bullshit. It claims that 'while diffusion works in high-dimensional datasets, it struggl
2. 122.▲
   by frotaur 3y ago

   Why should it be (1/2)*i improvement rate ? It could be 1/i ? I don't understand this argument.
3. 123.▲
   by frotaur 3y ago

   In the paper it is mentioned that the LLMs predicting the previous token are indeed pre-trained in this way, so it is not true that the difference is obvious.
4. 124.▲
   by frotaur 3y ago

   I don't think any AI or AI safety researcher imagines that using anything like "treating others as you would want to be treated by them" will solve the alignment problem
5. 125.▲
   by frotaur 3y ago

   No I think the point that was being made is that the "average person" idea is not that great if you have huge variance. If I have a uniform distribution from 0 to 1, average is '0.5', but its just as likely to get 0. or
6. 126.▲
   by frotaur 3y ago

   Are you sure the thermodynamic limit on water desalination is extremely small ? I was under the impression that the salt bonds dissolved by water are quite strong and thus require a non-negligible amount of energy to reverse (for instance,