Comment by cubefox | Hacker News Reader

Comment by cubefox | Hacker News Reader

> hyping the current grifter's fantasy of large language models as "artificial intelligence"

I really can't take him seriously after that remark. No, language models are of course artificial intelligence, they are systems so impressive that they were considered decades away before GPT-3 released a few years ago, or, to most, even before ChatGPT, just one year ago. They of course aren't precisely the classical SciFi intelligence that Stross and his colleagues had in mind in the past decades, but that speaks against them, not against the status of language models as AI.

By the way, a better SF writer, Stanisław Lem, did predict something similar to language models: Systems that would read novels of one author and then produce similar novels as output. Basically an LM without fine-tuning.

What Stross does is the classic goal-post moving, where anything more narrow than human intelligence is ultimately dismissed as not real AI. As an added bonus, he lampoons concerns about such "real" AI scaling to superintelligence as being totally speculative. Basically, either it's not real or it's just irresponsible speculation, simple, case closed!

I wish that I were good enough at communicating why I think that large language models are AI, but I just can't really argue against blanket statements like "it doesn't matter how it is implemented if it still appears intelligent".

People give me arguments like, it doesn't matter that LLMs don't really have any internal "thoughts" or ambitions, because each token is calculated solely from the previous tokens - that can be added on. You can chain LLMs together and do a bunch of this other stuff in order to implement that. "Internal thought and deliberation is just a feature you can implement on top of existing LLMs!"

Just... I get that AI won't necessarily have to work like how biological brains do, because intelligence is achievable in many different ways, but LLMs just... aren't it.

The people most against LLMs being described as AI seem to be the people who have a pre-existing notion for what AI is, and they point to all the things it can't do as reasons why it isn't actually AI. I've considered LLMs to be a piece of a greater and more complete AI in the same way our vocal cords are a piece of our internal thoughts manifesting in the real world so other people can hear them.

So an LLM is a great speaker, but it doesn't understand what it's saying. Does it need to? Do my vocal cords need to understand the stream of vibrations my brain asks for, or just respond? Understanding speech takes place at a different layer than merely repeating the "correct" speech.

Now all we need to do is attach the LLM output to a database for long term context storage, overlay a constant weight adjustment daemon that the LLM can trigger to control how it "learns", and tie the LLM temperature to /dev/random so it has mood swings like all of us meatbags do. THEN it will be a perfect human-like AI :P

> They of course aren't precisely the classical SciFi intelligence that Stross and his colleagues had in mind in the past decades

I've recently had the thought that there is an incredibly popular "AI" that is

1) Treated within universe as non-intelligent

2) Incredibly talented at understanding language and translating it directly into commands

3) Often compared unfavourably to actually intelligent AI

LLMs are the star-trek shipboard computer.

Good observation. Though to extend the analogy, LLMs may be capable of giving rise to true AI in the same way that the Enterprise's main computer could fairly trivially create sentient (and seemingly sapient) individuals like Dr. Moriarty, or the Doctor, despite not having that property itself.

4) Not allowed to make command decisions

I've not read Stanisław Lem (and have only seen the western adaption of Solaris, not the original), and I would therefore add an additional example, the three levels of AI used by Alastair Reynolds in Revelation Space (published in 2000): Alpha (sentient brain uploads); beta (non-sentient simulations based on what you do and say in your life, so kinda like GPT even though he didn't have that as an example to write from); and gamma (basically just normal programs).

Reply on news.ycombinator.com