I see a huge amount of confusion around terminology in discussions about Artificial Intelligence, so here’s my quick attempt to clear some of it up.

Artificial Intelligence is the broadest possible category. It includes everything from the chess opponent on the Atari to hypothetical superintelligent systems piloting spaceships in sci-fi. Both are forms of artificial intelligence - but drastically different.

That chess engine is an example of narrow AI: it may even be superhuman at chess, but it can’t do anything else. In contrast, the sci-fi systems like HAL 9000, JARVIS, Ava, Mother, Samantha, Skynet, or GERTY are imagined as generally intelligent - that is, capable of performing a wide range of cognitive tasks across domains. This is called Artificial General Intelligence (AGI).

One common misconception I keep running into is the claim that Large Language Models (LLMs) like ChatGPT are “not AI” or “not intelligent.” That’s simply false. The issue here is mostly about mismatched expectations. LLMs are not generally intelligent - but they are a form of narrow AI. They’re trained to do one thing very well: generate natural-sounding text based on patterns in language. And they do that with remarkable fluency.

What they’re not designed to do is give factual answers. That it often seems like they do is a side effect - a reflection of how much factual information was present in their training data. But fundamentally, they’re not knowledge databases - they’re statistical pattern machines trained to continue a given prompt with plausible text.

  • chemical_cutthroat@lemmy.world
    310·
    3 days ago

    When someone online claims that LLMs aren’t AI, my immediate response is to ask them to prove they are a real and intelligent life form. It turns out proving you are real is pretty damned hard when it boils down to it. LLMs may be narrow AI, but humans are pretty narrow in our thinking as well.

    I started a project back in January. It’s not ready for the public yet, but I’m planning for a early September release. Initially I don’t think it will be capable of much, but I’m going to be training it on various datasets in hopes that it is able to pick up on the basics fairly quickly. Over the next few years I’m aiming to train it on verbal communication and limited problem solving, as well as working on refining motor skills for interaction with its environment. After that, I’ll be handing it off regularly to professionals who have a lot more experience than me when it comes to training. Of course, I’ll still have my own input, but I’ll be relying a lot on the expertise of others for training data. It’s going to be a slow process, but my long term goal is a world wide release sometime in 2043, or maybe 2044, with some limited exposure before then. Of course, the training process never ends and new data is always becoming available, so I expect that to continue well beyond 2044.