back
5 comments
what might be better?

As the article points out, there is a definite need for something that can look beyond statistically parroting the rules of language.

Rain Man could speak English but he obviously had severe mental and emotional impairment as evidenced by the simple fact he thought everything cost about a hundred dollars.

We wouldn't allow Rain Man to legally drive a car. So why are we allowing AI with similar mental incapacity?

I have a feeling that comparison isn’t going to age well. Autism in humans (no matter how severe) isn’t really all that similar to stochastic parrotism in executables.
Yes, I'm thinking the lack of reasoning capacity in current AI executables is actually much more severe.

We need a standardized test of some sort that can quantify this fact.

Personally I’m of the opinion that the lack of reasoning capacity in LLMs is _total_… meaning an LLM doesn’t reason at all, in any sense, which is precisely why comparing its capacity for reasoning to an autistic savant is so very flawed and, if you think on it, rather inhumane. The character of Raymond (and the real human who inspired him) was absolutely capable of rich and meaningful reasoning, but due to a physical injury to the brain had sacrificed wide reasoning breadth the rest of us take for granted in favor of narrow reasoning depth we can scarcely imagine. The LLM isn’t wide or narrow, shallow or deep… it’s just got such an expansive model of the frequency in which the tokens in its training corpus appear that it can, frequently enough to be compelling, give us back a simulacrum of sufficient apparent reasonableness that the fact it contains absolutely zero novel reasoning whatsoever goes unnoticed by us (usually because we’re asking it a question we don’t really know the answer to, while also having a pretty poor internal model of just how many people contributed statistically identical questions and actually reasoned answers to those questions in the training corpus).

BTW, if we do ever develop a viable standardized test for actual reasoning capacity I’ll lay my money right now that not a single LLM (ever) passes it.

Ask AI about color of emergency brake in the Lufthansa Airplanes....