We warned Google that people might believe AI was sentient. Now it's happening
washingtonpost.com
washingtonpost.com
This is AI in general. As it exists today, there is no "intelligence". It's all database lookups, pattern matching and probabilities. Interesting and useful in some cases --- but still just as dumb as a hammer.
Until we can logically define "intelligence", it's not possible to program a binary computer to reproduce it. Logic playback is the only thing the box we call a "computer" can do. Claims to the contrary (such as this one involving "sentience") are pure cult fantasy.
"It's all database lookups, pattern matching and probabilities." These LLMs, and really all machine learning systems, are not database lookups at all, so I'll focus on the other half of your sentence. I would argue that human intelligence is just pattern matching and probabilities. Intelligence is nebulous to define, but I think the best concrete definition I've heard is that intelligence is a measure of the ability to integrate data in order to synthesize new data. Machine learning systems already do this, and thus have some level of intelligence, though perhaps not as much as humans yet simply because they don't integrate data as effectively as we do.
The example I always like to use is of a person who is born blind, and locked in a room their whole life with nothing but a bed and bland food to sustain them. But they also are given a huge library of audiobooks from which to learn about the world. They have no personal experience with anything they hear about, but from the language alone, they are able to obtain meaning through a recognition of patterns and an integration of given patterns into synthetic new ones.
In that analogy, I think it would be disingenuous to say this blind person is not intelligent -- or, to use your phrase, that they're "dumb as a hammer" -- and I think we'd all agree they are actually learning things through their audiobooks. And yet, they are doing the same thing these LLMs are: collecting linguistic data, finding patterns in it, and integrating information about those patterns in order to encode/synthesize new information.
Having other forms of data to incorporate definitely helps improve the effectiveness of the intelligence -- which is why multimodal language models are all the rage in bleeding-edge research -- but I don't think multimodality is necessary for intelligence.