Inspired by a recent talk from Richard Stallman.
From Slashdot:
Speaking about AI, Stallman warned that “nowadays, people often use the term artificial intelligence for things that aren’t intelligent at all…” He makes a point of calling large language models “generators” because “They generate text and they don’t understand really what that text means.” (And they also make mistakes “without batting a virtual eyelash. So you can’t trust anything that they generate.”) Stallman says “Every time you call them AI, you are endorsing the claim that they are intelligent and they’re not. So let’s let’s refuse to do that.”
Sometimes I think that even though we are in a “FuckAI” community, we’re still helping the “AI” companies by tacitly agreeing that their LLMs and image generators are in fact “AI” when they’re not. It’s similar to how the people saying “AI will destroy humanity” give an outsized aura to LLMs that they don’t deserve.
Personally I like the term “generators” and will make an effort to use it, but I’m curious to hear everyone else’s thoughts.
I disagree with this post and with Stallman.
LLMs are AI. What people are actually confused about is what AI is and what the difference between AI and AGI is.
There is no universal definition for AI, but multiple definitions which are mostly very similar: AI is the ability of a software system to perform tasks that typically would involve human intelligence like learning, problem solving, decision making, etc. Since the basic idea is basically that artificial intelligence imitates human intelligence, we would need a universal definition of human intelligence - which we don’t have.
Since this definition is rather broad, there is an additional classification: ANI, artificial narrow intelligence, or weak AI, is an intelligence inferior to human intelligence, which operates purely rule-based and for specific, narrow use cases. This is what LLMs, self-driving cars, assistants like Siri or Alexa fall into. AGI, artificial general intelligence, or strong AI, is an intelligence equal to or comparable to human intelligence, which operates autonomously, based on its perception and knowledge. It can transfer past knowledge to new situations, and learn. It’s a theoretical construct, that we have not achieved yet, and no one knows when or if we will even achieve that, and unfortunately also one of the first things people think about when AI is mentioned. ASI, artificial super intelligence, is basically an AGI but with an intelligence that is superior to a human in all aspects. It’s basically the apex predator of all AI, it’s better, smarter, faster in anything than a human could ever be. Even more theoretical.
Saying LLMs are not AI is plain wrong, and if our goal is a realistic, proper way of working with AI, we shouldn’t be doing the same as the tech bros.
Can you share the prompt you gave to ChatGPT to get this, I have questions and I want to cut out the middle man.
Feel free to ask your questions, I’ll gladly answer them. Before making stupid and smug claims, maybe you should’ve ran my post through literally any AI text detector before embarrassing yourself.
If I’m reading correctly it sounds like you do agree with Stallman’s main point that a casual distinction is needed, you just disagree on the word itself (“ANI” vs “generator”).
No, I think the distinction is already made and there are words for that. Adding additional terms like “generators” or “pretend intelligence” does not help in creating clarity. In my opinion, the current definitions/classifications are enough. I get Stallman’s point, and his definition of intelligence seems to be different from how I would define intelligence, which is probably the main disagreement.
I definitely would call a LLM intelligent. Even though it does not understand the context like a human could do, it is intelligent enough to create an answer that is correct. Doing this by basically pure stochastics is pretty intelligent in my books. My car’s driving assistant, even if it’s not fully self driving, is pretty damn intelligent and understands the situation I’m in, adapting speed, understanding signs, reacting to what other drivers do. I definitely would call that intelligent. Is it human-like intelligence? Absolutely not. But for this specific, narrow use-case it does work pretty damn good.
His main point seems to be breaking the hype, but I do not think that it will or can be achieved like that. This will not convince the tech bros or investors. People who are simply uninformed, will not understand an even more abstract concept.
In my opinion, we should educate people more on where the hype is actually coming from: NVIDIA. Personally, I hate Jensen Huang, but he’s been doing a terrific job as a CEO for NVIDIA, unfortunately. They’ve positioned themselves as a hardware supplier and infrastructure layer for the core component for AI, and are investing/partnering widely into AI providers, hyperscalers, other component suppliers in a circle of cashflow. Any investment they do, they get back multiplied, which also boosts all other related entities. The only thing that went “10x” as promised by AI is NVIDIA stock. They are bringing capex to a whole new level currently.
And that’s what we should be discussing more, instead of clinging to words. Every word that any company claims about AI should automatically be assumed to be a lie, especially for any AI claim from any hyperscaler, AI provider, hardware supplier, and especially-especially from NVIDIA. Every single claim they do directly relates to revenue. Every positive claim is revenue. Every negative word is loss. In this circle of money they are running - we’re talking about thousands of billions USD. People have done way worse, for way less money.