Comments by "Vitaly L" (@vitalyl1327) on "Francois Chollet - Why The Biggest AI Models Can't Solve Simple Puzzles" video.

  1. I don't see how LLMs (empowered with tools) cannot do ARC tests. You don't need to train models on puzzles, it's just general reasoning and pattern inference, something LLMs can already do well (even the very small ones, like Phi-3 medium). For example, LLMs write the best quality code in languages they do not know and did not see anything like in their training set. Just pair them with the right tools.
    1