<p>While writing after a tough week, I was editing my message on the fly a couple of times, as I made mistakes (like writing coffee instead of wine). I thought that unlike a human, an LLM doesn’t correct itself on the fly in the format of “oops, not that.”</p>
<p>This is related to the fact that the model predicts the next token each time but doesn’t influence the previous one. Yes, you can run the entire result through the model again, but that’s not quite what a human does.</p>
<p>It would be interesting to try to train/fine-tune a model that sometimes outputs the token “delete the last word” and “delete the last sentence.”</p>
<p>What do you think?</p>
<p>#ai #llm #gpt #finetunning</p>
<p><a href="https://t.me/+OvImEUmA7W5mYTRi">————————— Мысли Рвачева —————————</a></p>