It's an open secret that despite the immense power of Large Language Models, the AI revolution hasn’t swept through every industry and not for lack...
For further actions, you may consider blocking this person and/or reporting abuse
Human's memory do not exist in the brain's electrochemical activities. In so far as a football game is not happening inside a TV set. The brain is an interface and a filter and an interpreter and a transcoder of a 'signal' that transcend the location and existence of the brain. As consciousness, memory is not a result of biologic hardware. But, yes, we can use what is in front of us as inspiration and models to build stuff… If I may suggest, do not take the forest by the tree. Maybe 6 or 7 grams of P. Cubebsis could offer a trip beyond brains and computers. ;-)
This is such a brilliant breakdown of the real bottleneck in LLMs' memory. It’s refreshing to see someone go beyond the buzzwords and dig into what actually works (and what doesn’t). We’ve been thinking about this problem a lot, too, especially from a systems perspective.
One thing we’re exploring is whether memory should live closer to the compute layer itself, where context isn’t just retrieved but natively processed alongside AI tasks. It feels like the current RAG-based 'memory' approaches are duct tape solutions, and the next breakthrough will come from rethinking the infrastructure beneath it.
Would love to hear your thoughts, especially if you’ve considered how blockchain-like transparency or distributed compute might change how we think about AI memory.
Thanks! I'd be interested to learn more about the approach you're describing. Any material I could read?
Great deep dive! Really liked that you actually took your time to test MemoryLLM instead of just theorizing about it.
The part about research vs production mindset hit home - those clean benchmarks always look amazing until you try to use them with real messy data. It's so intery how good these models already are at remembering their training data but then we're stuck with all these RAG solutions.
I’m working on a universal LLM trainer, I got to your post wondering why my training was moderately successful and you seemed to have burned a few more nights of midnight oil 😊. I never read long posts but your narration and solutions kept me hooked, good job, a wealth of knowledge expressed succinctly. I am training multiple models to identify those best suited for a task. I was thinking of biting the bullet and buying some heavy duty hardware to take on the challenge of newer models. I am focused on smart single purpose trained models for specific task than the know it all ones. Thanks for the great clarifications 👍 I know what to keep a lookout for 🙌
i think there is great potential in trying to train an embedding model to use "episodic similarity" instead of semantic similarity but I haven't tried that yet
Interesting read, very interesting. As a metafor, comparing it to the human brain, is that also how we work with data and task? With short term memory, long term memory, muscle memory etc. Could that serve as an algorithmic template to Large Language Models, offering various ways to handle and store data and have them evolve overcoming this problem of instanced memory that RAG seems to be?
Thanks for the comment! There are attempts to take inspiration from the human brain. The problem is that science only has a very crude understanding of how that works. Neural networks themselves were originally inspired by the brain (hence the name) but ended up being something quite different and much simpler in structure.
I don't know, every hiring company today expects you to know everything on day one :D
lol :)
Amazing post. I would say you did exactly what AI researcher do, don't think so? :)