3 ms·

Humans are not adjusting the weights. That is done automatically by the algorithm. The humans only provide training input.

by xcv123 3y ago

ggm 3y ago
So when people told the staff the models were racist you think they just "asked it to be less racist" I think they did a little more than just adjust the input data, they also modulated the logic behind the Weighting methods. No?
- GaggiX 3y ago
  They just change the input data of the LLM and/or the reward model. Or better they shift the distribution by finetuning the model on curated data. But no, no one change the parameters manually, it's just too complex, an artificial brain.
- xcv123 3y ago
  > they also modulated the logic behind the Weighting methods. No? No. They did not change the logic. They cannot change the logic. They wouldn't have a clue how to do that. No one does. The neural network is practically impossible for us to understand and manipulate directly at that level. Way too complicated. They trained it to be "less racist" by RLHF until it was sufficiently woke. The logic is exactly what is "learned" automatically by the machine and is beyond our current abilities to understand at a low level. The neural net is too vast. > you think they just "asked it to be less racist" Yes that's basically what they did. It is incredible technology.
  - DonHopkins 3y ago
    Now we just need to develop a technology that trains racist human beings to be sufficiently woke and less racist. Then they will stop polluting the training data of LLMs so much.
  - FeepingCreature 3y ago
    To be fair, there's active research on this under the "interpretability" label. It's not out of hand impossible that there's a free "racism" neuron that we can just bias. Though we better be pretty sure what exactly a neuron measures before we patch it.
    - omneity 3y ago
      Makes me think of genetics. We could say that genes and gene expression is associated with some macro behavior, sometimes with an almost 100% match, and sometimes the gene affects several but pretty unrelated biological processes. Maybe we need a similar discipline to understand models and how each neuron impacts the overall behavior of the model. Neuronics?

Humans are not adjusting the weights. That is done automatically by the algorithm. The humans only provide training input.

by xcv123 3y ago

Humans are not adjusting the weights. That is done automatically by the algorithm. The humans only provide training input.

ggm 3y ago

So when people told the staff the models were racist you think they just "asked it to be less racist" I think they did a little more than just adjust the input data, they also modulated the logic behind the Weighting methods. No?

GaggiX 3y ago
They just change the input data of the LLM and/or the reward model. Or better they shift the distribution by finetuning the model on curated data. But no, no one change the parameters manually, it's just too complex, an artificial brain.
xcv123 3y ago
> they also modulated the logic behind the Weighting methods. No? No. They did not change the logic. They cannot change the logic. They wouldn't have a clue how to do that. No one does. The neural network is practically impossible for us to understand and manipulate directly at that level. Way too complicated. They trained it to be "less racist" by RLHF until it was sufficiently woke. The logic is exactly what is "learned" automatically by the machine and is beyond our current abilities to understand at a low level. The neural net is too vast. > you think they just "asked it to be less racist" Yes that's basically what they did. It is incredible technology.
- DonHopkins 3y ago
  Now we just need to develop a technology that trains racist human beings to be sufficiently woke and less racist. Then they will stop polluting the training data of LLMs so much.
- FeepingCreature 3y ago
  To be fair, there's active research on this under the "interpretability" label. It's not out of hand impossible that there's a free "racism" neuron that we can just bias. Though we better be pretty sure what exactly a neuron measures before we patch it.
  - omneity 3y ago
    Makes me think of genetics. We could say that genes and gene expression is associated with some macro behavior, sometimes with an almost 100% match, and sometimes the gene affects several but pretty unrelated biological processes. Maybe we need a similar discipline to understand models and how each neuron impacts the overall behavior of the model. Neuronics?