this post was submitted on 23 Dec 2023
189 points (86.8% liked)
Technology
59629 readers
2700 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Friendly reminder that your predictive text, while very compelling, is not alive.
It's not a mind.
Cyberpunk 2077 sorta explores this a bit.
There’s a vending machine that has a personality and talks to people walking by it. The quest chain basically has you and the vending machine chatting a bit and even giving the vending machine some advice on a person he has a crush on. You eventually become friends with this vending machine.
When it seems like it’s becoming more apparent it’s an AI and is developing sentience, it turns out the vending machine just has a really well-coded socializing program. He even admits as much when he’s about to be deactivated.
So, to reiterate what you said: predictive text and LLMs are not alive nor a mind.
I don't care, Brandon was real to me okay 😭
Which is why the Turing Test needs to be updated. These text models are getting really good at fooling people.
The Turing test isn't just that there exists some conversation you can have with a machine where you wouldn't know it's a machine. The Turing test is that you could spend an arbitrary amount of time talking to a machine and never be able to tell. ChatGPT doesn't come anywhere close to this, since there are many subjects where it quickly becomes clear that the model doesn't understand the meaning of the text it generates.
Exactly thank you for pointing this out. It also assumes that the tester would have knowledge of the wider context in which the test exists. GPT could probably fool someone from the middle ages, but that person wouldn't know anything about what it is they are testing for exactly.
I don't think most people will care, so long as their NPC interaction ends up compelling. We've been reading stories about people who don't exist for centuries, and that's stopped no one from sympathizing with them - and now there's a chance you could have an open conversation with them.
Like, I think alot of us assume that we care about the authors who write the character dialogs but I think most people actually choose not to know who is behind their favorite NPCs to preserve some sense that the NPC personality isn't manufactured.
Combine that with everyone becoming steadily more lonely over the years, and I think AI-generated NPC interactions are going to take escapism to another level.
Poem poem poem poem then the NPC start quoting Mein Kampf and killing all the cat wizards.
Lol, yeah. If generative AI text stays as shitty as it is now, then this whole discussion moot. Whether that will be the case has yet to be seen. What is an indisputable fact, though, is that right now is the worst that generative AI will ever be again. It's only able to improve from here.
That isn't actually true. With the rise in articles, posts and comments written by these algorithms, experts are warning about model collapse. Basically, the lack of decent human-written training data will destroy future generative AI before it can even start.
That's an interesting point. We are seeing a similar kind of issue with search engines losing effectiveness due to search engine optimization on websites.
So it is possible that generative AI will become enshittened.
Suppose you grew a small collection of brain cells and tied it into a CPU, would it be a mind then?
https://www.nature.com/articles/d41586-023-03975-7#:~:text=A%20system%20that%20integrates%20brain,hybrid%20machine%20can%20recognize%20voices.&text=Researchers%20have%20built%20a%20hybrid,tasks%20such%20as%20voice%20recognition.
If you cut out a tiny bit of someone's brain and then hooked it up to a cpu, would it be a mind? No, of course not, lol. Even if we got Biocomputers to work, we still wouldn't have any synthetic hardware even close to being strong or fast enough to actually create or even simulate a brain.
While it is not alive, whether it is a mind is not a clear cut. It can be called kind of a mind, a mind different of that of human.
What can't be a kind of mind to you?
Unless you want to call your predictive text on your keyboard a mind you really can't call an LLM a mind. It is nothing more than a linear progression from that. Mathematically proven to not show any form of emergent behavior.
No such thing has been "mathematically proven." The emergent behavior of ML models is their notable characteristic. The whole point is that their ability to do anything is emergent behavior.
Here's a white paper explicitly proving:
https://arxiv.org/abs/2304.15004
The field changes fast, I understand it is hard to keep up
Sure, if you define "emergent abilities" just so. It's obvious from context that this is not what I described.
Their paper uses industry standard definitions
Their paper uses terminology that makes sense in context. It's not a definition of "emergent behavior."
I do not think that it is “linear” progression. ANN by definition is nonlinear. Neither I think anything is “mathematically proven”. If I am wrong, please provide a link.
Sure thing: here's a white paper explicitly proving:
https://arxiv.org/abs/2304.15004
Thank you. This paper though does not state that there are no emergent abilities. It only states that one can introduce a metric with respect to which the emergent ability behaves smoothly and not threshold-like. While interesting, it only suggests that things like intelligence are smooth functions, but so what? Some other metrics show exponential or threshold dependence and whether the metric is right depends only how one will use it. And there is no law that emerging properties have to be threshold like. Quite the opposite - nearly all examples in physics that I know, the emergence appears gradually.
It is obvious that you do not know what either "mathematical proof" or "emergence" mean. Unfortunately, you are misrepresenting the facts.
I don't mean to criticize your religious (or philosophical) convictions. There is a reason people mostly try to keep faith and science separate.
Here's a white paper explicitly proving:
No emergent properties (illusory due to bad measures)
Predictable linear progress with model size
https://arxiv.org/abs/2304.15004
The field changes fast, I understand it is hard to keep up
As I said, you do not understand what these 2 terms mean. As such, you are incapable of understanding that paper.
Perhaps your native language is Italian, so here are links to the .it Wikipedia.
https://it.wikipedia.org/wiki/Comportamento_emergente
https://it.wikipedia.org/wiki/Dimostrazione_matematica
Emergence is the whole being greater than the sum of its parts. That's the original meaning of emergent properties, which is laid out in the first paragraph of the article. It's the scholarly usage as well, and what the claims of observed emergence are using as the base of their claim.
The article very explicitly demonstrated that only about 10% of any of the measures for LLMs displayed any emergence and that illusory emergence was the result of overly rigid metrics. Swapping to edit distance as an approximately close metric causes the sharp spikes to disappear for obvious reasons: no longer having a sharp yes/no allows for linear progression to reappear. It was always there, merely masked by flawed statistics.
If you can't be bothered to read here's a very easy to understand video by one of the authors: https://www.youtube.com/watch?v=ypKwNrmuuPM
Good. Now do you understand how you have misrepresented the paper?