Comments by "Vitaly L" (@vitalyl1327) on "3 Key Version Control Mistakes (HUGE STEP BACKWARDS)" video.

@vincentvogelaar6015 LLMs do have reasoning abilities. Just give them a Prolog, and let them write, test and debug proofs.
1
@vincentvogelaar6015 apparently you do not understand how to use LLMs. They're not any different from our own minds - we cannot reason either, unless we use tools, such as formal logic. So give LLMs their tools - give them an ability to write down reasoning step by step, to verify the reasoning using formal methods (as in - make them write down the steps as HoL proofs or Prolog predicates). I'm using LLMs for solving engineering problems, and reasoning is a crucial part of it. Even very small models (like Phi-3) are perfectly capable of reasoning on a level beyond the capacity of an average engineer, when given the right tools and proper sandbox to test the ideas in (akin to our imagination). Also, LLMs perform the best when reasoning about things that were not in the training set. E.g., they write much better code in languages they've never seen - because they're forced to do it slowly, verifying every step, instead of churning out answers instinctively.
1
@vincentvogelaar6015 apparently you do not understand how to use LLMs. They're not any different from our own minds - we cannot reason either, unless we use tools, such as formal logic. So give LLMs their tools - give them an ability to write down reasoning step by step, to verify the reasoning using formal methods (as in - make them write down the steps as HoL proofs or Prolog predicates). Give them a sandbox to debug the proofs, just like you do with any other code. Provide a critical loop to make sure they did not miss anything from the formulation of the problem when translating it to a proof. I'm using LLMs for solving engineering problems, and reasoning is a crucial part of it. Even very small models (like Phi-3) are perfectly capable of reasoning on a level beyond the capacity of an average engineer, when given the right tools and proper sandbox to test the ideas in (akin to our imagination). Also, LLMs perform the best when reasoning about things that were not in the training set. E.g., they write much better code in languages they've never seen - because they're forced to do it slowly, verifying every step, instead of churning out answers instinctively.
1
@vincentvogelaar6015 youtube keeps deleting my answer for some reason. I'll try to make it shorter: in order to reason, LLMs must use tools, just like we do. Writing things down, visualising them, applying formal logic, you know, formally - and it all works really well with even the very small LLMs (like Phi-3). And LLMs perform the best on knowledge that was not in the training set, it's forcing them to slow down and reason in very small verifiable steps instead of generating an "intuitive" answer.
1
@vincentvogelaar6015 how is it not a reasoning, if this is exactly what humans do? With the result at least not worse, but often much better than human reasoning. You just give this autonomous agent loop with its toolbox a single, vaguely worded goal, and after some time and a lot of very small steps it will create a solution. Fully autonomously. Right now I'm using LLMs to design mixed signal circuits and mechanical parts this way. Just give them the goal, and they do the work. Including finding the datasheets for the required parts, creating and running simulations for things they're not certain about, etc.
1
@vincentvogelaar6015 see the Voyager paper for a rudimentary example of such a self-improvement. No need for further model training - there are more efficient ways of storing the accumulated knowledge.
1