I’m willing to debate the potentials of AI again once they manage to do that without those “benchmarks” getting special attention in the training data.

You sound like those guys who doomed AI, because a single neuron wasn’t able to solve the XOR problem. Guess what, build a network out of neurons and the problem is solved.

What potentials are you talking about? The potentials are tremendous. There are a plethora of algorithms, theoretic knowledge and practical applications where AI really shines and proves its potential. Just because LLMs currently still lack several capabilities, this doesn’t mean that some future developments can’t improve on that and this by maybe even not being a contemporary LLM. LLMs are just one thing in the wide field of AI. They can do really cool stuff. This points towards further potential in that area. And if it’s not LLMs, then possibly other types of AI architectures.