I don’t think they’re useless and I was pleasantly surprised to see how effective training on expressions of uncertainty was (got rid of 2/3 of confabulations and didn’t cause a big drop in answering when the model is correct) but empirically it doesn’t seem to be enough to fix things

September 21, 2025 - 17:25 UTC

1
0
4
There isn't even anything to fix; correctness can never be more than an accidental product of the system's use
2
0
6
Yea if only someone had been making correctness machines for the last 100 years or so we could lean on to help ai somehow
0
0
1
Exactly, it is by definition rolling dice for every output. It assigns probability to the next pool of tokens (words, word fragments, symbols), then literally randomly selects based on the assigned probabilities. It is a bullshitting machine FIRST AND FOREMOST. Any other perspective is incorrect.
1
0
3