The idea of a "compressed 21st Century" is a bit strange. An analogy would be maybe that the invention of the microscope led to a "compressed 500 years" that allowed advances in medicine in 250 years that would have taken 500 years. It's not really the case because the microscope was necessary for the advances, and they wouldn't have happened without it. The counterfactual is wrong. Just like manipulation of big biological data requires AI. It's just another tool that provides new affordances. But anyway, he's still very upbeat about all this "Powerful AI" that is on the way (again, in 5-10 years), compared to Apple's new conclusions: https://arxiv.org/abs/2410.05229 (has Apple not released serious AI because they're not happy with how present day AI functions (or does not function)? 100% agree with your comments Ben about the environment/ecological blindspot of almost everyone working it tech. But I guess that's ok because the AI coming in 5-10 years will 'solve' this.
Thanks for that Matt - you are a stickler for precision! Personally I'm happy with the "compressed century" concept more as a subjective than objective framing. Due to the fact that now we can viscerally *feel* the acceleration happening in far less than a human lifespan - whereas 1 century, 2 centuries, five centuries... are all much the same as "after I die". Therefore the "objective" counterfactual isn't really the point I take away... more that previous models we had of the rate of progress are likely substantially wrong.
(And also: yes, Apple's paper has caused a lot of the expected anti-AI noise in *Gary Marcus circles* but any controversy seems to come down to which interpretation of "reasoning" is adopted and whether you think a human brain "reasoning" does anything fundamentally different to an LLM... not a lot to see here for me ...)
I think the key issue for me is one of error correction, which LLMs seem to be terrible at (even 99% is no good if you're making 100 critical decisions a minute), and actually structurally incapable of solving. It comes back to the underlying architecture, and this makes me think of Jeff Hawkins and his original 'hierarchical temporal memory' analysis. It's the top down constant prediction, expectation, detection, correction, iteration, that allows humans to eg catch a ball (to use one of Hawkins early examples) and this applies across all cognition (yes human brains are fundamentally different to LLMs - though LLM-like components will probably be part of the overall architecture, and we might be able to learn a lot about human cognition from analysis of LLMs). Need architecture additional to/other than standard deep neural nets to make the error rate anything near reliable enough for critical deployments (eg finance, aviation, governance). I also think that a key issue is that the training material for code is quite good (because people tend to upload working code, or buggy code with the tag 'this isn't working, help'. So the learning material is high quality for code. But for human language there are all kinds of examples of falsehoods and fallacies, divergent beliefs, etc all over the web, so the learning material is problematic. LLMs will be much better coders than writers.
I definitely rate Jeff Hawkins' model as the closest (intuitively, anyway) to what's actually going on in a brain. He's gone quiet since writing his book... wonder what he's working on right now?
The idea of a "compressed 21st Century" is a bit strange. An analogy would be maybe that the invention of the microscope led to a "compressed 500 years" that allowed advances in medicine in 250 years that would have taken 500 years. It's not really the case because the microscope was necessary for the advances, and they wouldn't have happened without it. The counterfactual is wrong. Just like manipulation of big biological data requires AI. It's just another tool that provides new affordances. But anyway, he's still very upbeat about all this "Powerful AI" that is on the way (again, in 5-10 years), compared to Apple's new conclusions: https://arxiv.org/abs/2410.05229 (has Apple not released serious AI because they're not happy with how present day AI functions (or does not function)? 100% agree with your comments Ben about the environment/ecological blindspot of almost everyone working it tech. But I guess that's ok because the AI coming in 5-10 years will 'solve' this.
Thanks for that Matt - you are a stickler for precision! Personally I'm happy with the "compressed century" concept more as a subjective than objective framing. Due to the fact that now we can viscerally *feel* the acceleration happening in far less than a human lifespan - whereas 1 century, 2 centuries, five centuries... are all much the same as "after I die". Therefore the "objective" counterfactual isn't really the point I take away... more that previous models we had of the rate of progress are likely substantially wrong.
(And also: yes, Apple's paper has caused a lot of the expected anti-AI noise in *Gary Marcus circles* but any controversy seems to come down to which interpretation of "reasoning" is adopted and whether you think a human brain "reasoning" does anything fundamentally different to an LLM... not a lot to see here for me ...)
I think the key issue for me is one of error correction, which LLMs seem to be terrible at (even 99% is no good if you're making 100 critical decisions a minute), and actually structurally incapable of solving. It comes back to the underlying architecture, and this makes me think of Jeff Hawkins and his original 'hierarchical temporal memory' analysis. It's the top down constant prediction, expectation, detection, correction, iteration, that allows humans to eg catch a ball (to use one of Hawkins early examples) and this applies across all cognition (yes human brains are fundamentally different to LLMs - though LLM-like components will probably be part of the overall architecture, and we might be able to learn a lot about human cognition from analysis of LLMs). Need architecture additional to/other than standard deep neural nets to make the error rate anything near reliable enough for critical deployments (eg finance, aviation, governance). I also think that a key issue is that the training material for code is quite good (because people tend to upload working code, or buggy code with the tag 'this isn't working, help'. So the learning material is high quality for code. But for human language there are all kinds of examples of falsehoods and fallacies, divergent beliefs, etc all over the web, so the learning material is problematic. LLMs will be much better coders than writers.
I definitely rate Jeff Hawkins' model as the closest (intuitively, anyway) to what's actually going on in a brain. He's gone quiet since writing his book... wonder what he's working on right now?