Those “reasoning models” are my favorite. It’s basically the equivalent of adding another pass through the generator with the additional prompt “now sprinkle in some text that makes it look like you are thinking about each part of your answer”.
Yup, pretty much. And the field is full of red herring terms, so they can mislead you into believing otherwise: “hallucination”, “semantic” supplementation, “reasoning” models, large “language” model…
BradleyUffner@lemmy.world 5 weeks ago
BananaIsABerry@lemmy.zip 5 weeks ago
LLM creators: *feeds an algorithm millions of lines of text
Some dude on the internet: “language”
lvxferre@mander.xyz 5 weeks ago
I use those quotation marks because IMO they’re better described as large token models than large language models. They have rather good morphology and syntax, but once you look at the higher layers (semantics and specially pragmatics) they drop the ball really hard. Even if those layers are way more important than the lower ones.
For a rough analogy, it’s like a taxidermised cat - some layers (the skin and fur) are practically identical to the real thing, but it’s missing what makes a cat a cat, you know? It’s still useful if you want some creepy deco, but don’t expect the taxidermised critter to ruin your furniture or to use your belly as sleeping pad.
bear@lemmy.blahaj.zone 5 weeks ago
I’m going to be very disappointed if Elon’s AI wins.
lvxferre@mander.xyz 5 weeks ago
Do you want my guess? The current “fight” will go on, until the AI bubble bursts. None of the current large token models will survive; they’ll be simply ditched as “unprofitable”. Instead you’ll see a bunch of smaller models popping up, for more focused tasks, being advertised as something else than AI (perhaps as a “neural network solution” or similar).
So Grok, Gemini, GPT, they’re all going the way of the dodo.
That’s just my guess though. It could be wrong.
snooggums@piefed.world 5 weeks ago
Small focused learning models and other forms of AI have been used for decades.
The current bubble is just trying to make LLMs do literally everything including accurately answering questions despite their core design including randomization to appear more like a human.
lvxferre@mander.xyz 5 weeks ago
Yes, but I think the ones you’ll see past the bubble burst will be a bit different. For example, incorporating the current attempts of natural language processing, even if in a simplified way.