4 ms·
still no chance. Just think of five blind persons and one one-eyed talking about a movie they saw yesterday. While it may make up perfect sense the guys talking
by theGeatZhopa 2y ago
still no chance. Just think of five blind persons and one one-eyed talking about a movie they saw yesterday. While it may make up perfect sense the guys talking about, its still not possible to find what the one-eyed said.
So you take each one's answer and take parts out it synthesize/aligning them all to a perfect sensed prompt. Fed again back into the LLMs, you'll maaaayyyy find whats the proper answer, by finding contradictions in the returns.
Just assume, 5 blind & 1 1-eyed' tell you about the movie. You'll get 6 different answers. Now you take them, mix them and feed them back into LLM. What will happen?
- may be the 1-eyed will correct you / correct the false parts. So you might know "ok, that one is obviously been in the movie. Because he correct me and the others don't"
- may be all of the blinds will correct you and the one-eyed (lets call him cotton' eye joe) dont do it. How you find out which one is correct?
What is actually needed is a stress-maker, a "cortex"-LLM and a few of other ML-types for the domain expertise (like math, physics, ...). The cortex is the main LLM that will be prompted. Cortex utilizes domain experts for answering. Stress-maker always keep the cortex busy in the times, when its not used by a prompter, by synthesizing "questions" that has been seen but answered incorrectly before.
That will simulate our brain, where we have an stress-maker that keep our brain always in thinking. And the other experts like tasting, sensing, hearing, seeing.. etc. But I don't know this is a good approach to solve the problem right now.
Whats needed now is an LLM that can utilize the other LLMs. A multi-translator that understands the syntax and the data of other models. If that is put into existence, than its also possible to build a multi-modal-agent that can fuzz its way into different knowledge-bases and file-formats and what ever. This would be incredible.