We just released François Chollet keynote talk from #AGI24. He challenges LLM hype, discusses ARC-AGI benchmark, and proposes merging deep learning with program synthesis for true AGI progress.

@fchollet I propose merging deep learning and counterfactual regret minimisation

I mean everyone has always known, at least those of us actually who have been working on this for longer than ChatGPT, that machine learning needs an expert systems level grounding for real generality and flexible intelligence…we’re just closer now than we’ve ever been…and it may just be a simple ontology placed over the reasoning engine of an LLM to do it🤔🤷🏼‍♂️🥳🦾

There is a new form of mental illness, a type of delusional disorder related to magical thinking, where the sufferers are unable to distinguish the difference between LLM output and actual human style cognition, even when the differences are clearly pointed out to them and they are given unambiguous examples.

@fchollet 🤦

@fchollet FTR, Chollet has plagiarized my work and that is not his idea but mine like his previous ideas.

@fchollet Absolutely excellent talk, he had already uploaded the original recording on the arc agi yt channel and I've watched it several times. It's the best formalisation of intelligence I've seen anyone give and very relevant to the "are LLMs reasoning" discussion

@fchollet ha! he looks like talking about Symbols and Symboliad in my framework of atoms of meaning. I go deep around Symbol's Omniload and externsion. Observer pattern analogy. LLMs need a broader arc of generativity and data (kaleidoscope is not intricate) .

@fchollet Bringing us the likes of Monsieur Chollet and Herr Schmidhuber week after week. Ambassador, you're really spoiling us!

@fchollet This is worth a watch.