[birdsongs]

> LLM's are notoriously bad at this. The noise to signal ratio is unacceptably high

I could go either way on the future of this, but if you take the argument that we're still early days, this may not hold. They're notoriously bad at this so far.

We could still be in the PC DOS 3.X era in this timeline. Wait until we hit the Windows 3.1, or 95 equivalent. Personally, I have seen shocking improvements in the past 3 months with the latest models.

[kombookcha]

Personally I strongly doubt it. Since the nature of LLM's does not allow them semantic content or context, I believe it is inherently a tool unsuited for this task. As far as I can tell, it's a limitation of the technology itself, not of the amount of power behind it.

Either way, being able to generate or compress loads of text very quickly with no understanding of the contents simply is not the bottleneck of information transfer between human beings.

[birdsongs]

Yeah, definitely more skeptical for communication pipelines.

But for coding, the latest models are able to read my codebase for context, understand my question, and implement a solution with nuance, using existing structures and paradigms. It hasn't missed since January.

One of them even said: "As an embedded engineer, you will appreciate that ...". I had never told it that was my title, it is nowhere in my soul.md or codebase. It just inferred that I, the user, was one. Based on the arm toolchain and code.

It was a bit creepy, tbh. They can definitely infer context to some degree.

[lelanthran]

> We could still be in the PC DOS 3.X era in this timeline. Wait until we hit the Windows 3.1, or 95 equivalent. Personally, I have seen shocking improvements in the past 3 months with the latest models.

While we're speculating, here's mine: we're in the Windows 7 phase of AI.

IOW, everything from this point on might be better tech, but is going to be worse in practice.

[mcny]

I would like to see the day when the context size is in gigabytes or tens of billions of tokens, not RAG or whatever, actual context.

[riskable]

Context size helps some things but generally speaking, it just slows everything down. Instead of huge contexts, what we need is actual reasoning.

I predict that in the next two to five years we're going to see a breakthrough in AI that doesn't involve LLMs but makes them 10x more effective at reasoning and completely eliminates the hallucination problem.

We currently have "high thinking" models that double and triple-check their own output and we call that "reasoning" but that's not really what it's doing. It's just passing its own output through itself a few times and hoping that it catches mistakes. It kind of works, but it's very slow and takes a lot more resources.

What we need instead is a reasoning model that can be called upon to perform logic-based tests on LLM output or even better, before the output is generated (if that's even possible—not sure if it is).

My guess is that it'll end up something like a "logic-trained" model instead of a "shitloads of raw data trained" model. Imagine a couple terabytes of truth statements like, "rabbits are mammals" and "mammals have mammary glands." Then, whenever the LLM wants to generate output suggesting someone put rocks on pizza, it fails the internal truth check, "rocks are not edible by humans" or even better, "rocks are not suitable as a pizza topping" which it had placed into the training data set as a result of regression testing.

Over time, such a "logic model" would grow and grow—just like a human mind—until it did a pretty good job at reasoning.

[aakresearch]

Upvoted, as it basically 99% matches my own thinking. Very well said. But I, personally, would not predict a breakthrough in this direction in the next 2-5 years, as there is no pathway from current LLM tech to "true reasoning". In my mental model LLM operates in "raster space" with "linguistic tokens" being "rasterization units". For "true reasoning" an AI entity has to operate fluently in "vector space", so to speak. LLM can somewhat simulate "reasoning" to a limited degree, and even that it only does with brute force - massive CPU/GPU/RAM resources, enormous amount of training data and giant working contexts. And still, that "simulation" is incomplete and unverifiable.

I would argue that the research needed to enable such "vector operation" is nowhere near the stage to come to fruition in the next decade. So, my prediction is, maybe, 20-50 years for this to happen, if not more.

[2snakes]

Wasn’t this idea the basic premise of coq? Why didn’t it work?

[lelanthran]

> I would like to see the day when the context size is in gigabytes or tens of billions of tokens, not RAG or whatever, actual context.

Might not make a difference. I believe we are already at the point of negative returns - doubling context from 800k tokens to 1600k tokens loses a larger percentage of context than halving it from 800k tokens to 400k tokens.

[butlike]

First impressions are everything. It's going to be hard to claw back good will without a complete branding change. But... where do you go from 'AI'???

[Izkata]

There's many things that used to be called AI, but as their shortcomings became known we started dropping them from the AI bucket and referring to them by a more specific name: expert systems, machine learning, etc. Decades later plenty of people never learned this and those things don't pop into mind with "AI" so LLMs were able to take over the term.

Given time I could see this happening again.