If you're an LLM trying to symbolically emulate human behavior, especially when coding, you're going to see git commits, where people write a bunch of code and then go, oh wait, I should fix that!
From the perspective of an outside observer, making mistakes, and then correcting the mistakes, is how you code.
In fact, an outside observer might think that this is how you teach people to code. You make a mistake, then you say, oh, here's the mistake I made, and here's why, and then you fix it.
So I think true emergent behavior will be demonstrated, when an LLM says "Hey wait a minute! They're actually making real mistakes, and they're actually correcting, it's them not on purpose?!", and then stops hallucinating, and stops making mistakes as it codes.
The same is true of bug trackers. You submit a bug request and you often have a patch. Then people discuss it. From a MLM's perspective, you're supposed to write poor code, or code that's not perfect, and then have a conversation.
If I had the resources to train a large-scale LLM, I would clone, for example, GitHub. I would then remove everything where people are fixing broken code.
The outcome would be very interesting.
I've been looking at ways to make enhanced long-term memory stores for LLMs, and there's lots of problems with shifting symbolic relationships if you do it wrong, but definitely once there's true long-term memory, and adaptive behavio,r I think that's the only way you're going to get true real emergent behavior.
I think when people are irritated by the "hallucination" aspect of LLMs, they are often running into something of a slightly different nature. I mean, firstly, there seems to be a higher-than-normal ratio of these "little mistakes". And secondly, there are some pretty odd ones - e.g. in my team, Claude regularly just straight up makes up Jira ticket numbers, and then refers to them with high confidence. I guess what I'm saying is that a human would probably not just make up an id, and run with it (they may be off by one, or mix two up, etc). In my opinion, these can be successfully treated, but I guess you can never fully eliminate the tendency that irks people.
I could have used any series of NLP examples, the point is this CoT behavior only emerges when you get to a certain scale (and training style, presumably).
> They're Made out of Meat
https://web.mit.edu/people/dpolicar/writing/prose/text/think...
To me the heart of the "next token predictor" is that the distributions are static. You can manipulate what you feed into it through context (and a lot of interesting engineering has been applied there through CoT and other techniques to manipulate the prompt). But these models as implemented will never be able try things and learn from mistakes or adapt. They are a set of weights frozen in time. A set of distributions derived from the original data that created them.
And there's nothing inherently stopping labs from continuously fine-tuning the weights after every new invocation. It's just a difficult (not to mention expensive) software and ML problem.
What prevents continuous fine-tuning from what I understand is catastrophic forgetting. You can do things like RLHF which are built to minimize the damage but that is more about bringing out capabilities of the base model than incorporating new knowledge (at least from my understanding, I am obviously not a researcher at a lab).
My broader point is just that there's nothing inherent to the structure of LLMs that stops them from updating their weights and continuously learning from environmental feedback in the way humans do, and there's already solid templates for how they could push even further in that direction.
But as an assessment of the current state, I agree with you, LLMs lag humans severely in ability to self-update.
1. Technical cost of updating the mode.
2. Inability to trust every user's "truth".
3. Ability of AGENT-HARNESSES to learn with the help of the human user.
So agents learn, LLM already knows everything it will ever know, and ESPECIALLY it has already learned how to understand human language.
No 3. above means there is no danger of the LLM getting corrupted. But the agents running on user's machine learn on behalf of that user who shares the machine with them.
"LLM" is a branded model as a product. Of course it could be anything, as long as it fulfills the product category.
But we live in reality, we can only look at what models are out there and we see that they don't do any of those things and yet we're supposed to act as if these models already do.
RNNs and Mamaba do not update their weights, but you could hypothetically scale the internal state to be as big as Fable's and GPT 6's parameters.