this post was submitted on 03 Jun 2024
1292 points (96.4% liked)
Technology
60078 readers
3301 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 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I have to disagree.
Frankly, LLMs (which are based on neural networks) seem a Hell of a lot closer to how actual brains work than "classical AI" (which basically boils down to a gigantic pile of
if
statements) does.I guess I could agree that LLMs are undeserving of the term "AI", but only in the sense that nothing we've made so far is deserving of it.
"seem" is the critical word there. Interacting with an LLM they do seem to be pretty clever.
I'm not talking about interacting with it. I'm talking about how it's implemented, from my perspective as a computer scientist.
Let me say it more concretely: if even shitty expert systems, which are literally just flowcharts implemented in procedural code, are considered "AI" -- and historically speaking, they are -- then the bar is really fucking low. LLMs, which at least make an effort to kinda resemble the structure of biological intelligence, are certainly way, way above it.
I'm actually sad that the state of AI deserves the hate it gets. Neural networks are so sick, just going through the example of detecting a diagonal on a 2x2 grid was like magic to me. And they made me second guess simulation theory for quite a while lmao
Tangentially, blockchain was a similar phenomenon for me. Or at least trust networks. One idea was to just throw away Certificate Authorities. Basically federate all the things, and this was before we knew about the fediverse. It gets all the hate because of crypto, but it's cool tech. The CA thing would probably lead to a bad place too, though.
The fuck?
Let's agree to disagree then. An LLM has no notion of semantics, it's just outputting the most likely word to follow up to what it's already written and the user's input.
On the contrary, expert systems from back in the 90s for, say, predicting the atomic structure of an element, work like a human brain on steroids. It features an arbitrary large search tree that the software knows how to iterarively prune according to a well known set of chemical rules. We do the same when analyzing a set of options.
Debugging "current" AI models, on the other hand, is impossible because all we're doing is prescripting a composition of functions and forcing it to minimize a loss function. That's all we're doing. How can you currently tell that a certain model is going to work? Unless the mathematical theory ever catches up with the technology, we'll never know until we execute the code.