OK, current large language models are trained by a huge chunk of text and for every token displayed the model has to predict the next token based on the loss function and all the other statistics out there that points to the most likely token, but the model predicts by just trillions of parameters. Those are numeric numbers represented for a large chunk of text trained tomorrow. That’s why the model can hallucinate because just the numeric numbers the primary numeric numbers which doesn’t represent what the model actually knows, but it will be represents the the general data the general perimeter that the output will output the next most likely token based on that data so what I’m saying is that the huge chunk of text retrained or fine to know everything like that, I converted to that configurations of the number of parameters, right but the configuration of thae configuration of the number of primes do not mean anything. They are just like if we didn’t know anything about the date before they were just be showing numbers by numbers that are generated because of that data but it doesn’t represent anything. Does it explain what the one knows doesn’t do any of that that’s why they hallucinations that’s why they.

My proposal is that why don’t we train tomorrow? Given the test the model would know and understand what it has, and then infer from the text so that when the user asked about it, we can explain it based on the data that it receives and understand and if the user ask something that isn’t on the data the model since it already knows the data and always receive the data it has we solidified the data right solidify the data and the model explain to the use of the best on my current data. I do not have that have that and I can refer for my daughter. I thought it might be something like that. This isn’t that so like a huge huge change in the way we train tomorrow you know we don’t give it like a huge Texan give it like a and I’ll put like a conjugation configurations of numbers instead we give it the data and it understands and then it predicts it answers based on the data it doesn’t answer based on the configuration configuration of parameters note it answers based on the data based on the actual texture data