Will Depue argues that AGI is the underlying learning algorithm rather than static trained model weights

@theojaffee right

@theojaffee Nothing inert is intelligent

@Bayesian0_0 the main post from him about it being the "learning algorithm, not the weights" is a bit misleading, which is probably why he clarified it in the comment

I think your analogy goes a step too far

The brain is the AGI, not what gave rise to it

Genes/environment play a role in that, but they are not themselves the intelligence. The genes are like the AGI's blueprint and the environment, well I just the environment in which it learns

In your analogy you would also equate the researchers who built the AGI to AGI

@scaling01 Isn’t this like saying human intelligence is the genes that encode brain learning and the environment you grow up in, not the synaptic connections / the brain itself? Which seems wrong

@theojaffee "Learning algorithm" is not enough for AGI. Weights are memories and LLMs need the abilities to "CRUD" it's weights.

@theojaffee

@theojaffee damn, read it from him, amazing insight.

@theojaffee Ding! Ding! Ding!

We have a winner!

The OpenAI Jalepeno hardware only targets inference, not training. As their frontier model unifies weights, inference and context.

@theojaffee that’s my dad telling my mom that their son has a learning problem… and my mom kept feeding me

my simplified view is like this: - the system that trains the model is the AGI, not the model (which is what he meant with his first post) - the brain is the system and the neocortex is the model

the models weights themselves are not AGI, the neocortex is not AGI

but the weight thing is a bit muddy, because I think the "AGI" / brain is also modifying how it learns, not just the neocortex

@theojaffee damn he is private now, why did he do that lol

@theojaffee People who believe in the bitter lesson would argue this is false

@theojaffee LLMs are the engine, not the car.

@haidongly @theojaffee Weights are not just memories. Its also abstracted out processing (ie, eigenwords)

They do no local (or global of course) consistency. No mining of implied knowledge

Its a weird sort of frozen A{not quite G}I

Its an interesting time

@scaling01 the structured loop is more important than the underlying model it uses as the engine beyond a certain point of model capability