Comment by cs702
21 days ago
The problem is even more fundamental: Today's models stop learning once they're deployed to production.
There's pretraining, training, and finetuning, during which model parameters are updated.
Then there's inference, during which the model is frozen. "In-context learning" doesn't update the model.
We need models that keep on learning (updating their parameters) forever, online, all the time.
Why is learning an appropriate metaphor for changing weights but not for context? There are certainly major differences in what they are good or bad at and especially how much data you can feed them this way effectively. They both have plenty of properties we wish the other had. But they are both ways to take an artifact that behaves as if it doesn't know something and produce an artifact that behaves as if it does.
I've learned how to solve a Rubik's cube before, and forgot almost immediately.
I'm not personally fond of metaphors to human intelligence now that we are getting a better understanding of the specific strengths and weaknesses these models have. But if we're gonna use metaphors I don't see how context isn't a type of learning.
I suppose ultimately, the external behaviour of the system is what matters. You can see the LLM as the system, on a low level, or even the entire organisation of e.g. OpenAI at a high level.
If it's the former: Yeah, I'd argue they don't "learn" much (!) past inference. I'd find it hard to argue context isn't learning at all. It's just pretty limited in how much can be learned post inference.
If you look at the entire organisation, there's clearly learning, even if relatively slow with humans in the loop. They test, they analyse usage data, and they retrain based on that. That's not a system that works without humans, but it's a system that I would argue genuinely learns. Can we build a version of that that "learns" faster and without any human input? Not sure, but doesn't seem entirely impossible.
Do either of these systems "learn like a human"? Dunno, probably not really. Artificial neural networks aren't all that much like our brains, they're just inspired by them. Does it really matter beyond philosophical discussions?
I don't find it too valuable to get obsessed with the terms. Borrowed terminology is always a bit off. Doesn't mean it's not meaningful in the right context.
To stretch the human analogy, it's short term memory that's completely disconnected from long term memory.
The models currently have anteretrograde amnesia.
It’s not very good in context, for one thing. Context isn’t that big, and RAG is clumsy. Working with an LLM agent is like working with someone who can’t form new long term memories. You have to get them up to speed from scratch every time. You can accelerate this by putting important stuff into the context, but that slows things down and can’t handle very much stuff.
The article does demonstrate how bad it is in context.
Context has a lot of big advantages over training though, too, it's not one-sided. Upfront cost and time are the big obvious ones, but context also works better than training on small amounts of data, and it's easier to delete or modify.
Like even for a big product like Claude Code from someone that controls the model, although I'm sure they do a lot of training to make the product better, they're not gonna just rely entirely on training and go with a nearly blank system prompt.
You got this exactly backwards.
"I'm not fond of metaphors to human intelligence".
You're assuming that learning during inference is something specific to humans and that the suggestion is to add human elements into the model that are missing.
That isn't the case at all. The training process is already entirely human specific by way of training on human data. You're already special casing the model as hard as possible.
Human DNA doesn't contain all the information that fully describes the human brain, including the memories stored within it. Human DNA only contains the blue prints for a general purpose distributed element known as neurons and these building blocks are shared by basically any animal with a nervous system.
This means if you want to get away from humans you will have to build a model architecture that is more general and more capable of doing anything imaginable than the current model architectures.
Context is not suitable for learning because it wasn't built for that purpose. The entire point of transformers is that you specify a sequence and the model learns on the entire sequence. This means that any in-context learning you want to perform must be inside the training distribution, which is a different way of saying that it was just pretraining after all.
The fact the DNA doesn't store all connections in the brain doesn't mean that enormous parts of the brain, and by extension, behaviour aren't specified in the DNA. Tons of animals have innate knowledge encoded in their DNA, humans among them.
I don't think it's specific to humans at all, I just think the properties of learning are different in humans than they are in training an LLM, and injecting context is different still. I'd rather talk about the exact properties than bemoan that context isn't learning. We should just talk about the specific things we see as problems.
Models gain information from context but probably not knowledge and definitely not wisdom.
> We need models that keep on learning (updating their parameters) forever, online, all the time.
Do we need that? Today's models are already capable in lots of areas. Sure, they don't match up to what the uberhypers are talking up, but technology seldom does. Doesn't mean what's there already cannot be used in a better way, if they could stop jamming it into everything everywhere.
Continuous learningin current models will lead to catastrophic forgetting.
will catastrophic forgetting still occur if a fraction of the update sentences are the original training corpus?
is the real issue actually catastrophic forgetting or overfitting?
nothing prevents users from continuing the learning as they use a model
3 replies →
How long will it take someone to poison such a model by teaching it wrong things?
Even humans fall for propaganda repeated over and over .
The current non-learning model is unintentionally right up there with the “immutable system” and “infrastructure as code” philosophy.
> How long will it take someone to poison such a model by teaching it wrong things?
TayTweets was a decade ago.
> The current non-learning model is unintentionally right up there with the “immutable system” and “infrastructure as code” philosophy.
As long as training material remains the proprietary secret sauce, the average user doesn’t already see or benefit from that - it’s all a promise and a black box to us.
I'm not sure if you want models perpetually updating weights. You might run into undesirable scenarios.
If done right, one step closer to actual AGI.
That is the end goal after all, but all the potential VCs seem to forget that almost every conceivable outcome of real AGI involves the current economic system falling to pieces.
Which is sorta weird. It is like if VCs in Old Regime france started funding the revolution.
I think VCs end up in one of four categories
1. They're too stupid to understand what they're truly funding.
2. They understand but believe they can control it for their benefit, basically want to "rule the world" like any cartoon villain.
3. They understand but are optimists and believe AGI will be a benevolent construct that will bring us to post scarcity society. There are a lot of rich / entrepreneurs that still believe they are working to make the world a better place.. (one SaaS at a time but alas, they believe it)
4. They don't believe that AGI is close or even possible
If it makes the models smarter, someone will do it.
From any individual, up to entire countries, not participating doesn't do anything except ensure you don't have a card to play when it happens.
There is a very strong element of the principles of nature and life (as in survival, not nightclubs or hobbies) happening here that can't be shamed away.
The resource feedback for AI progress effort is immense (and it doesn't matter how much is earned today vs. forward looking investment). Very few things ever have that level of relentless force behind them. And even beyond the business need, keeping up is rapidly becoming a security issue for everyone.
3 replies →
Yes the planet got destroyed. But for a beautiful moment in time we created a lot of value for shareholders.
And for your comparison, they did fund the American revolution which on its turn was one of the sparks for the French revolution (or was that exactly the point you were making?)
2 replies →
1. Progress is unstoppable. Refusing to fund it won't make it disappear.
2. Most VCs are normal people that just want a bigger slice of pie, not necessarily a bigger share of the pie. See the fixed pie fallacy.
Our brains, which are organic neural networks, are constantly updating themselves. We call this phenomenon "neuroplasticity."
If we want AI models that are always learning, we'll need the equivalent of neuroplasticity for artificial neural networks.
Not saying it will be easy or straightforward. There's still a lot we don't know!
I wasn't explicit about this in my initial comment, but I don't think you can equate more forward passes to neuroplasticity. Because, for one, simply, we (humans) also /prune/. And... Similar to RL which just overwrites the policy, pushing new weights is in a similar camp. You don't have the previous state anymore. But we as humans with our neuroplasticity do know the previous states even after we've "updated our weights".
How would you keep controls - safety restrictions - Ip restrictions etc with that, though? the companies selling models right now probably want to keep those fairly tight.
3 replies →
Tay the chatbot says hi from 2017.
How about we just put them to bed once in a while?
Please elaborate on this one
1 reply →
it is interesting
Please elaborate
Doesn't necessarily need to be online. As long as:
1. there's a way to take many transcripts of inference over a period, and convert/distil them together into an incremental-update training dataset (for memory, not for RLHF), that a model can be fine-tuned on as an offline batch process every day/week, such that a new version of the model can come out daily/weekly that hard-remembers everything you told it; and
2. in-context learning + external memory improves to the point that a model with the appropriate in-context "soft memories", behaves indistinguishably from a model that has had its weights updated to hard-remember the same info (at least when limited to the scope of the small amounts of memories that can be built up within a single day/week);
...then you get the same effect.
Why is this an interesting model? Because, at least to my understanding, this is already how organic brains work!
There's nothing to suggest that animals — even humans — are neuroplastic on a continuous basis. Rather, our short-term memory is seemingly stored as electrochemical "state" in our neurons (much like an LLM's context is "state", but more RNN "a two-neuron cycle makes a flip-flop"-y); and our actual physical synaptic connectivity only changes during "memory reconsolidation", a process that mostly occurs during REM sleep.
And indeed, we see the same exact problem in humans and other animals, where when we stay awake too long without REM sleep, our "soft memory" state buffer reaches capacity, and we become forgetful, both in the sense of not being able to immediately recall some of the things that happened to us since we last slept; and in the sense of later failing to persist some of the experiences we had since we last slept, when we do finally sleep. But this model also "works well enough" to be indistinguishable from remembering everything... in the limited scope of our being able to get a decent amount of REM sleep every night.
It 100% needs to be online. Imagine you're trying to think about a new tabletop puzzle, and every time a puzzle piece leaves your direct field of view, you no longer know about that puzzle piece.
You can try to keep all of the puzzle pieces within your direct field of view, but that divides your focus. You can hack that and make your field of view incredibly large, but that can potentially distort your sense of the relationships between things, their physical and cognitive magnitude. Bigger context isn't the answer, there's a missing fundamental structure and function to the overall architecture.
What you need is memory, that works when you process and consume information, at the moment of consumption. If you meet a new person, you immediately memorize their face. If you enter a room, it's instantly learned and mapped in your mind. Without that, every time you blinked after meeting someone new, it'd be a total surprise to see what they looked like. You might never learn to recognize and remember faces at all. Or puzzle pieces. Or whatever the lack of online learning kept you from recognizing the value of persistent, instant integration into an existing world model.
You can identify problems like this for any modality, including text, audio, tactile feedback, and so on. You absolutely, 100% need online, continuous learning in order to effectively deal with information at a human level for all the domains of competence that extend to generalizing out of distribution.
It's probably not the last problem that needs solving before AGI, but it is definitely one of them, and there might only be a handful left.
Mammals instantly, upon perceiving a novel environment, map it, without even having to consciously make the effort. Our brains operate in a continuous, plastic mode, for certain things. Not only that, it can be adapted to abstractions, and many of those automatic, reflexive functions evolved to handle navigation and such allow us to simulate the future and predict risk and reward over multiple arbitrary degrees of abstraction, sometimes in real time.
https://www.nobelprize.org/uploads/2018/06/may-britt-moser-l...
That's not how training works - adjusting model weights to memorize a single data item is not going to fly.
Model weights store abilities, not facts - generally.
Unless the fact is very widely used and widely known, with a ton of context around it.
The model can learn the day JFK died because there are millions of sparse examples of how that information exists in the world, but when you're working on a problem, you might have 1 concern to 'memorize'.
That's going to be something different than adjusting model weights as we understand them today.
LLMs are not mammals either, it's helpful analogy in terms of 'what a human might find useful' but not necessary in the context of actual llm architecture.
The fact is - we don't have memory sorted out architecturally - it's either 'context or weights' and that's that.
Also critically: Humans do not remember the details of the face. Not remotely. They're able to associate it with a person and name 'if they see it again' - but that's different than some kind of excellent recall. Ask them to describe features in detail and maybe we can't do it.
You can see in this instance, this may be related to kind of 'soft lookup' aka associating an input with other bits of information which 'rise to the fore' as possibly useful.
But overall, yes, it's fair to take the position that we'll have to 'learn from context in some way'.
8 replies →
Models like Claude have been trained to update and reference memory for Claude Code (agent loops) independently and as a part of compacting context. Current models have been trained to keep learning after being deployed.
yes but that's a very unsatisfactory definition of memory.
> models that keep on learning
These will just drown in their own data, the real task is consolidating and pruning learned information. So, basically they need to 'sleep' from time to time. However, it's hard to sort out irrelevant information without a filter. Our brains have learned over Milenial to filter because survival in an environment gives purpose.
Current models do not care whether they survive or not. They lack grounded relevance.
Maybe we should give next-generation models fundamental meta goals like self-preservation and the ability to learn and adapt to serve these goals.
If we want to surrender our agency to a more computationally powerful "consciousness", I can't see a better path towards that than this (other than old school theism).
> meta goals like self-preservation
Ah, so Skynet or similar.
Is this correct? My assumption is that all the data collected during usage is part of the RLHF loop of LLM providers. Assumption is based on information from books like empire of ai which specifically mention intent of AI providers to train/tune their models further based on usage feedback (eg: whenever I say the model is wrong in its response, thats a human feedback which gets fed back into improving the model).
... for the next training run, sure (ie. for ChatGPT 5.1 -> 5.2 "upgrade"). For the current model? No.
> We need models that keep on learning (updating their parameters) forever, online, all the time.
Yeah, that's the guaranteed way to get MechaHilter in your latent space.
If the feedback loop is fast enough I think it would finally kill the internet (in the 'dead internet theory' sense). Perhaps it's better for everyone though.
Many are working on this, as well as in-latent-space communication across models. Because we can’t understand that, by the time we notice MechaHitler it’ll be too late.
I don't understand why that's on the critical path. I'd rather a frozen Ramanujan (+ temporary working memory through context) than a midwit capable of learning.
I wish that agents could “sleep and consolidate” like humans do.
We need models that are smarter than humans. So far, the cost of an AI query + training is dwarfing the effort it would take to teach an intelligent human how to do a task. We are dumping an incredibly amount of money/effort into making AI do stuff when it's still not competitive with humans, because dumbass people are controlling investment. The stock market is not a replacement for competent investment. The fact people buy meme coins shows how fucked we are.
Deceiving people is not a sustainable business model, but it is the most prominent one in the US right now. Lie to the public, sell them stuff that's bad for them at too high of a price, get rich quick, then act confused when your economy collapses because the victims of your grift can't spend anymore.
Thanks for repeating what the author explained.
I think they can do in-context learning.