Our brains only developed their abstract reasoning capabilities after we already possessed a cognitive ‘platform’ for learning in a very physical context. Or rather, our cognitive capabilities for things that may seem totally extraneous to reasoning in fact developed in tandem with it. Think, fine motor control over our fingers and opposable thumbs, and the corresponding development of spatial and physical reasoning and hierarchical planning, that allows us to analyse a problem and build a tool with our hands to solve it.
The ‘bitter lesson’ [0] seems to be that we humans are not very good at designing the algorithmic machinery for cognition—better we let the machine discover its own mechanisms. Take the case of AlphaGo, the performance of which greatly improved when the human data was thrown away.
So, perhaps there is a pathway to artificial reasoning that shortcuts past many functions of an artificial brain, as you put it, but it also looks quite like we’re not ourselves going to be able to architect it, and that reasoning has not emerged from the current LLM paradigm of digesting the written knowledge of all humanity.
Where human general-purpose abstract reasoning naturally arose from concrete, goal-directed interaction with the physical world, perhaps we need to recreate that environment for machines to learn to learn and reason themselves.
That need not necessarily be physically embodied, which would surely be heavily constrained in numbers of learning trials. Quickly searching turned up this [1] example of reinforcement learning to walk in a sim, which is then transferred to physical.
I wonder how far the approach can go... Could agents learn to talk to each other? Looks like there’s some recent research in that direction too [2].
Our brains only developed their abstract reasoning capabilities after we already possessed a cognitive ‘platform’ for learning in a very physical context. Or rather, our cognitive capabilities for things that may seem totally extraneous to reasoning in fact developed in tandem with it. Think, fine motor control over our fingers and opposable thumbs, and the corresponding development of spatial and physical reasoning and hierarchical planning, that allows us to analyse a problem and build a tool with our hands to solve it.
The ‘bitter lesson’ [0] seems to be that we humans are not very good at designing the algorithmic machinery for cognition—better we let the machine discover its own mechanisms. Take the case of AlphaGo, the performance of which greatly improved when the human data was thrown away.
So, perhaps there is a pathway to artificial reasoning that shortcuts past many functions of an artificial brain, as you put it, but it also looks quite like we’re not ourselves going to be able to architect it, and that reasoning has not emerged from the current LLM paradigm of digesting the written knowledge of all humanity.
Where human general-purpose abstract reasoning naturally arose from concrete, goal-directed interaction with the physical world, perhaps we need to recreate that environment for machines to learn to learn and reason themselves.
That need not necessarily be physically embodied, which would surely be heavily constrained in numbers of learning trials. Quickly searching turned up this [1] example of reinforcement learning to walk in a sim, which is then transferred to physical.
I wonder how far the approach can go... Could agents learn to talk to each other? Looks like there’s some recent research in that direction too [2].
[0]: https://www.cs.utexas.edu/~eunsol/courses/data/bitter_lesson...
[1]: https://www.figure.ai/news/reinforcement-learning-walking
[2]: https://www.ifaamas.org/Proceedings/aamas2024/pdfs/p2725.pdf