There’s a fundamental difference between AI hallucinations and humans believing something which is incorrect.

With an AI, it’s a problem of statistics. The context window biases the model in just the right way to produce an answer which is incorrect. But, to the model, the answer was correct. That’s to say, it didn’t just fuck up its calculations… no, it arrived at the correct, albeit incorrect, solution.

That’s not the same as what a human experiences in its “hallucinations.” Not even close. Often, a human can sense that something is off (to say the least).

A human can work out their own error. An LLM can not. That’s not because an LLM isn’t good enough to do so. It’s because the LLM is missing something important. Once you add that thing into the architecture, I’d argue it’s de-facto no longer an LLM. Because, by that point, they’re leveraging new science for new behaviors in a much more complex system which overcomes the universal scale limitations of LLMs.

That’s my logic in a nutshell. LLMs might “lead” to AGI simply for the fact that it makes obvious we still haven’t achieved AGI, and so people would ideally start looking toward other approaches.

To me, an LLM is perhaps one subsystem in what should be a pretty large ecosystem of tooling. AGI would likely be the whole system, not just the LLM and really the LLM doesn’t deserve the credit (imo).