I recently read, maybe by you, that machines were being allowed to look at a style of problem and make a LoRA with it, before they tackle the actual problem. It was part of how to teach them to solve pattern recognition problems. The 'what happens next' ones that we get as part of psych evals.
And I thought that made sense. When I'm confronted with a problem that I've not seen before, I hunt around in my head for what is similar to the problem. Ways I can break the problem I'm given into chunks I can do. None of this relies on me coming up with solutions from scratch. All of it relies on what I've solved before.
And then when that fails, I start testing ideas.
So it seems to me that given LLM's just aren't thinkers yet, to test them as if they are is a waste of time.
Given an LLM the ability to make itself a set of LoRAs over time (if that's an efficient way to store a solution style) so that they have a library of experience to draw from. Just like we do.
We have to give them time to chew on the problem. Try things. Break the problem into bits. Etc.
I think that's the path to AGI because as I've said many times, an 8 yr old human has not been taught everything on the planet, and they are a quite capable GI. This strongly indicates it is structure, not knowledge, that will make an AGI.
The average crow is a better problem solver at the moment.
So my extra thought is, why can't we use them to do what we know they are good at? Why is everyone so keen to build a single model that is perfect at everything? Are they trying to build a genie outside of a bottle that they can just throw any task at?
Isn't it safer for us if we know the machines we use have some fairly hard limits. Than in certain areas they are dumb as posts?