> If you appeared in a puff of smoke before the authors of that paper, just after publication — a few months before half of them cleaved from OpenAI to form Anthropic — and carried with you a laptop linked through time to the big models of 2026, what would their appraisal be ? There’s no doubt in my mind they would say: Wow, we really did it ! This is obviously AGI!
I really don't think this would be the reaction. I'd say they would (or should) look at the systems we have now and see a very clear path between where they were then and where we are now, with all the positives _and negatives_. We still get hallucinations. We still get misalignment, if anything as capabilities have improved so has the potential for damage when things go wrong. It's pretty clear to me that late 2025 models are just better versions of what we had in 2021.
That's not to say they're not more useful, more valuable, they absolutely are! But that's all about product integrations, speed, and turning up the dial on inference compute. They're still fundamentally the same things.
The next big step forward, the thing that LLMs are obviously missing, is memory. The fact we're messing around with context windows, attention across the context space, chat lookup and fact saving features, etc, are all patches over the fact that LLMs can't remember anything in the way that humans (or pretty much any animal) can. It's clear that we need a paradigm shift on memory to unlock the next level of performance.
We have LLM memory, it's a training data from which the model was initially programmed. To allow adding or changing LLM memory, we would need to retrain model completely or partially. And that is not realistic any time soon. All other attempts at LLM memory would be just an obscure hack of splitting context window into parts and feeding input from different files. Literally nothing would change if you input half of the query from one file, half from another called "memory.txt" or if you just input whole query from a single file twice as big.
> It's clear that we need a paradigm shift on memory to unlock the next level of performance.
I think this is on point to the next phase of LLMs or a different neural network architecture that improves on top of them, alongside continual learning.
Adding memory capabilities would mostly benefit local "reasoning" models than online ones as you would be saving tokens to do more tasks, than generating more tokens to use more "skills" or tools. (Unless you pay more for memory capabilities to Anthropic or OpenAI).
It's kind of why you see LLMs being unable to play certain games or doing hundreds of visual tasks very quickly without adding lots of harnesses and tools or giving it a pre-defined map to help it understand the visual setting.
As I said before [0], the easiest way to understand the memory limitations with LLMs is Claude Playing Pokemon with it struggling with basic tasks that a 5 year old can learn continuously.
These are all very much in the same category of hacks that I mentioned.
A cat doesn't know its way around a house when it's born, but it also doesn't have to flick through markdown files to find its way around. A child can touch a hot stove once and be neurotic about touching hot things for the rest of their life, without having to read flash cards each morning or think for a few minutes about "what do I know about stoves" every time they're in the kitchen.
Call them a "hack" all you want, they seem to work. What's particularly intesting is how claude has been trained on skills, so it doesn't need to be taught how to use a skill, so that's been baked into it.
I'm not claiming they don't work in some sense, but as a user you have to be fairly deeply aware of how they work, context engineering is A Thing, you have to tell LLMs to remember stuff, etc.
We're hacking around the fact that the models don't learn in normal use. That's in no way controversial.
A model that continuously learnt would not need the same sort of context engineering, external memory databases, etc.
You speak the truth but looking back, what I reacted to is
> It's clear that we need a paradigm shift on memory to unlock the next level of performance.
and my take is that we might not need to get there to get the next level of performance, based on how well the latest models are able to utilize these hacks of a memory feature. On top of that, Claude was specifically RLHF'd to have the skills concept, so it's good with those. We disagree. Let's let time see who ends up being right.
I really don't think this would be the reaction. I'd say they would (or should) look at the systems we have now and see a very clear path between where they were then and where we are now, with all the positives _and negatives_. We still get hallucinations. We still get misalignment, if anything as capabilities have improved so has the potential for damage when things go wrong. It's pretty clear to me that late 2025 models are just better versions of what we had in 2021.
That's not to say they're not more useful, more valuable, they absolutely are! But that's all about product integrations, speed, and turning up the dial on inference compute. They're still fundamentally the same things.
The next big step forward, the thing that LLMs are obviously missing, is memory. The fact we're messing around with context windows, attention across the context space, chat lookup and fact saving features, etc, are all patches over the fact that LLMs can't remember anything in the way that humans (or pretty much any animal) can. It's clear that we need a paradigm shift on memory to unlock the next level of performance.