Hacker News Reader

Comment by rglover | Hacker News Reader

Possibly, but that's going to require 100% consistent, accurate outputs (tricky as that's not the nature of LLMs).

Otherwise, you'll have a lot of systems dependent on these orchestrators creating hard-to-debug mistakes up and down the pipeline. With software, you can reach a state where it does what you tell it to without having to worry if some model adjustment or API change is going to break the output.

If they solve that, then yes. Otherwise, what I personally expect is a lot of businesses rushing into implementing "agents" only to backpedal later when they start to have negative material effects on bottom lines.

View on HN

Its inevitable. You can argue about what's possible right now, but I'm not looking at it from that angle. I think these issues will be solved with time

Why? Past progress =/= equal rate of future progress.

That belief is at odds with the mechanics of how LLMs work. It's not a question of more effort/investment/compute/whatever, it's just a reality of how the underlying systems work (non-deterministic). If you can find a way to make the context window on the scale of the human brain, you may be able to mostly mitigate this.

People want us to be at "Her" levels of AI, but we're at a far earlier stage. We can fake certain aspects of that (using TTS), but blindly trusting an AI to run everything is going to be a big mistake in the short-term. And in order for the inevitability of what you describe to take place, the predecessor(s) to that have to work in a way that doesn't scare people and businesses away.

The plowing of money and hype into the current forms of AI (not to mention the gaslighting about their ability) makes me think the real inevitability is a meltdown in the next 5-10 years which leads to AI-hesitancy on a mass scale.

Have you tried o1 pro? I find people that are making these assertions are not deeply using the models on a daily basis. With each new release, I can see the increase of capability, and can do things. I have written software in the last year that is at a level of complexity beyond my skill set. I have 15 years of SWE experience, most at FAANG. You just arent close enough to the metal to see what's coming. It's not about what we have now, its about scaling and a reliable march of model improvements. The code has been cracked, given sufficient data, anything can be learned. Neural networks are generalized learners

Yes, I use LLMs every day. Primarily for coding (a mix of Claude and OAI). I was trying to implement a simple CSS optimization step to my JS framework's build system last night and both kept hallucinating to the point (literally inventing non-existent APIs and config patterns) where I gave up and just did it by hand w/ Google and browsing docs.

The problem with your "close to the metal" assertion is that this has been parroted about every iteration of LLMs thus far. They've certainly gotten better (impressively so), but again, it doesn't matter. By their very nature (whether today or ten years from now), they're a big risk at the business level which is ultimately where the rubber has to hit the road.

Yeah I don't think we're going to come closer to a real AGI until we manage to make a model that can actually understand and think. An LLM sounds smart but it just picks the most likely response from the echoes of a billion human voices. I'm sure we'll get there but not with this tech. I'm pretty sure even be OpenAI said this with their 5 steps to AGI, LLMs were only step 1. And probably the part that will do the talking in the final AI but not the thinking.

At the moment people are so wooed by the confidence of current LLMs that they forget that there's all sorts of types of AI models. I think the key is going to be to have them work together, each doing the part they're good at.

> An LLM sounds smart but it just picks the most likely response from the echoes of a billion human voices.

This is where reasoning models come in. Train models on many logical statements then give them enough time to produce a chain of thoughts that’s indistinguishable from “understanding” and “thinking”.

I’m not sure why this leap is so hard for some people to make.

I personally don't think that will go very far. It's just a way of extracting a little bit more out of a technology that's the wrong one for the purpose.

We just are not on your level of genius to understand these things.

So obviously completely full of shit.

Broadly I agree with your position, but:

> If you can find a way to make the context window on the scale of the human brain, you may be able to mostly mitigate this.

Human brains have a much smaller context window than AI do. We can't pay attention to the last 128,000 concepts that filtered past our sensory systems — our conscious considerations are for about seven things.

There's a lot of stuff that we don't yet understand well enough to reproduce with AI, but context length is the wrong criticism for these models.

> context length is the wrong criticism for these models

You're right. What I'm getting at is the overall speed, efficiency, and accuracy of the storage, retrieval, and processing capability of the human brain.

It's kinda crazy that it can run on a few slices of bread when LLMs need kilowatts of power to write a simple paragraph :)

They are using infinity compute and can’t do simple notifications. How will changing the architecture slightly or ingesting more data change that?

Reply on news.ycombinator.com