@nazokiyoubinbou @futurebird
Yeah. What I’m saying is that “agentic AI” isn’t just an LLM making guesses; it has a deterministic framework _around_ the LLM that _does_ apply logic (deterministic programs) to test what the LLM spits out. The LLM hallucinates a response, the framework fact-checks it & feeds back errors.
The LLM is not reasoning. The system seems to do something like what we do when we think we’re reasoning.
@Red_Shirt_no2 @futurebird I mean, even then it's not really. I suppose external tooling may have a check mechanism or two that applies computer logic, but they still lack understanding. The logic is a simple gate "is x true? Then y" type of thing. (Yes it gets more complex, but in other words it's not understanding, it's measurements.)
As for the "reasoning" mechanism, the LLM still is just applying statistical probabilities, just it has a training to produce a sort of double check to increase the likelihood of producing a more correct statistical probability. This helps reduce the effects of the randomization of the token production that produces what people call hallucinations without having to lower the temperature too much, bit it still lacks understanding or actual logic.
@Red_Shirt_no2 @futurebird As a side note, you say it looks like our thought process, but I guess you haven't actually seen the "thinking"? Local models that do it don't hide it from the user entirely. (It goes inside an HTML collapse.) The "thought" process goes something like this:
The user has asked us so-and-so. We should produce a response that fits within said criteria. First we should do x, then we should do y.
It's not really "thinking" of course. It's trained to produce an output that says "the user asked us so-and-so, we should so-and-so." That's just the training forcing that specific response first. But this raises the statistical chance of being able to produce a "wait, that isn't correct, I should say y instead of x" inside the hidden section.
@Red_Shirt_no2 @nazokiyoubinbou
I thought "agentic" just meant that you gave the LLM usernames and passwords to things like bank accounts and heating systems and instructed it to make a macro with a series of commands to be executed using those user accounts.
That is using a LLM to "do things" rather than just make text.
The additional systems that make this possible don't really change the fundamental model all that much, they just vastly expand the potential impact of any mistakes.
@futurebird @Red_Shirt_no2 Yeah, basically it's just sending specific outputs to tools designed to do the actual things. The tools still lack understanding in the same way notepad doesn't understand what you're doing when you type, just that it must produce text onto the screen as you type. They have specific, very narrow tasks that the API calls do.
They just train it to output the correct things to those APIs (most of the time...)