Re:Peter Norvig critically reviews AlphaCode's code quality
by visarga on Friday December 16, 2022 @05:19PM
(#34021329)
It all boils down to having some sort of embodiment, or a way to verify. For code it would suffice to let the model generate and execute code, and learn from errors. Give it enough "experience" with code execution and it will learn on its own, like AlphaGo. Generate more data and retrain the models a few times.
Re:Peter Norvig critically reviews AlphaCode's code quality
by kqr on Saturday December 17, 2022 @05:02AM
(#34026455)
> If you give a crawling baby a small cliff but make a bridge out of glass, the baby will refuse to cross it, because it's limited understanding prevents it from knowing that the glass is safe to crawl on and it won't fall.
Not just that -- it's enough to paint a grid on a flat floor using perspective manipulation to make it look like a steep drop.
...on the other hand, I once went onto an open-wall elevator in a VR game, which then started going up. Although I know I had just moved sideways on solid floor and there was still solid floor next to me, it was quite frightening to stand at the edge of it. I had to force myself really hard to tap my foot on the floor outside the elevator that didn't look like it existed.
There's a bunch of famous research that shows a baby and toddlers have basic understanding of physics. If you give a crawling baby a small cliff but make a bridge out of glass, the baby will refuse to cross it, because it's limited understanding prevents it from knowing that the glass is safe to crawl on and it won't fall.
In contrast older humans, even those with a fear of heights, are able to recognize that properly strong glass bridges are perfectly safe, and they won't fall through them just because they can see through them.
What changes when you go from one to the next? Is it just more data fed into the feedback machine, or does the brain build entirely new circuits and pathways and systems to process this more complicated modeling of the world and info it gets?
Everything about machine learning just assumes it's the first, with no actual science to support it, and further claims that neural nets with back-propagation are fully able to model that system, even though we have no idea how the brain corrects errors in it's modeling and a single neuron is WAY more powerful than a small section of a neural network.
These are literally the same mistakes made all the time in the AI field. The field of AI made all these same claims of human levels of intelligence back when the hot new thing was "expert systems" where the plan was, surely if you make enough if/else statements, you can model a human level intelligence. When that proved dumb, we got an AI winter.
There are serious open questions about neural networks and current ML that the community just flat out ignores and handwaves away, usually pretending that they are philosophy questions when they aren't. "Can a giant neural network exactly model what the human brain does" is not a philosophy question.