Hacker News new | ask | show | jobs
by zaking17 901 days ago
I like the process that goes into these "imagine the architecture of AGI" articles. It's all hypothetical, but it's really fun.

But it's a missed opportunity if you don't embed LLMs in some of the core modules -- and highlight where they excel. LLMs aren't identical to any part of the human brain, but they do a remarkable job of emulating elements of human cognition: language, obviously, but also many types of reasoning and idea exploration.

Where LLMs fail is in lookup, memory, and learning. But we've all seen how easy it is to extend them with RAG architectures.

My personal, non-scientific prediction for the basic modules of AGI are:

- LLMs to do basic reasoning

- a scheduling system that runs planning and execution tasks

- sensory events that can kick off reasoning, but with clever filters and shortcuts

- short term memory to augment and improve reasoning

- tools (calculators etc.) for common tasks

- a flexible and well _designed_ memory system -- much iteration required to get this right, and i don't see a lot of work being done on it, which is interesting

- finally, a truly general intelligence would have the capability to mutate many of the above elements based on learning (LLM weights, scheduling parameters, sensory filters, and memory configurations). But not everything needs to be mutable. many elements of human cognition are probably immutable as well.

3 comments

I like to think we could quickly create a next-level AI (maybe AGI?) if we simply model it on the Pixar movie "Inside Out". The little characters inside the girl's brain are different LLMs with different biases. They follow a kind of script that adapts to the current environment. They converse with each other and suggest to the girl what she should do or say.

I'd try the idea myself, but I have a job. :-)

This sounds a lot like the mixture-of-experts architecture, and the current best-performing language models (GPT-4, mixtral-8x7b) already use this architecture.

So congratulations, you win!

That's not really how MoEs work. They never directly interact with eachother. There is one manager type model that takes a prompt, directs token inference to 1 or more models, chooses the best response, and continues. The analogy would be closer to a "swarm of agents". (There are a handful of names for this approach, I think swarm is catching on the most)
one important thing you left out - the ability to reproduce and thus "evolve" naturally, and at scale, to essentially keep improving its own brain to the point it outpaces current human researchers in self-improvement. If not reproduce, maybe reincarnate itself in version 2.0, 3.0, etc...
Yeah, I guess I was heading in that direction with the last point. Earth organisms have a separation between lifetime learning (brain modification) and genetic evolution, but, for AGI, these could be combined into one, or further separated into three or more methods of goal-directed modification.
I think LLMs did not exist or barely existed at the time he wrote this.
Makes sense!