4 ms·

> you can have your voice as a thoroughly vetted example, and then be able to use that to finetune LLMs You have piqued my interest. Can you explain in more d

• randall 2y ago
  So the goal initially is going to be to create a "constitution." Basically a document that outlines your voice, or whatever. From there, you can create examples for every sentence and classify them... ie "this is a good example" or "this is almost a good example." Then, after you have a good enough enough number of human versions (idk this number yet), you can use those to create dozens more examples, refine, etc... then at some point you've thoroughly described a document enough that there's no more delta. Then you can use the document to train / finetune / infer. The effects are that the "instincts" of the model (if on the training / weights) or the "thoughts" of the model (at inference time) are closer to a well vetted baseline.