HeheheI'm waiting for your mail for SiriusGPT
. BTW, it is going to be an extra 2-3 days because something dawned on me recently. You know how you and I were discussing how I need to figure out the best way to combine the two model results? Well, I figured out a really cool way to do so, but it is going to take quite a bit of work... at least 2-3 days.I just sent the following to a MT member, and instead of typing everything again, here is how SiriusLLM and SiriusGPT are going to work...
It is going to be a couple more days than expected... as I was trying to figure out how to handle the interaction between the two models, it dawned on me that we should send the request to the first LLM, and if the confidence is say less than 75%, then to send the request to the second model to see if it can achieve a higher confidence. Then the more I thought about it, I was like "why stop there? Why not send it to 3-4 models until we find one that has the correct verdict with a high confidence"." Then I also realized that if the first 3-4 models are unable to achieve a high confidence, then we can send it to a super expensive, but amazingly accurate model that I found a month or so ago while testing tons of different models, and hopefully it will achieve a high confidence verdict. It is 10-20x the price of the other models, but it is dead on with the verdicts. And they just released a brand new model that is supposed to make the previous model look stupid
So then after figured out that we can just iterate the models until it finds a suitable one, I kind of got carried away and came up with several more methods / techniques to implement into Sirius. I even filed a provisional patent on it... which took all day today.
It is actually even more complicated then that, but that is the general idea.
So I have my work cut out for me... let's just hope nothing else "dawns" on me
.
