A reading from 2018-10-18: Still, I think, very wise on actually useful really-existing "AI" as very high-dimensional very big-data non-parametric regression & classification...
> 2. On making people conform to machines, rather than the other way around: That's what I was trying to get at in the last paragraph. In a different context I'd have expanded on it, and been much more explicit.
> 3. On generalization ("transfer") across games: I think my remarks largely hold. There's been progress, but it's stuff like this <https://proceedings.neurips.cc/paper_files/paper/2022/hash/b2cac94f82928a85055987d9fd44753f-Abstract-Conference.html>, where they train on 41 related videogames and generalize to 6 other related videogames. The key point to my mind is that, to quote their own conclusions "Our results are based largely on performance in the Atari suite, where _action and observation spaces are aligned across different games_" (my emphasis). I will admit that I have not followed stuff on game-play as closely since then as I probably should be, so if Tolley (or anyone else) has pointers to papers, I'd be appreciative.
> 4. I was actually excessively dismissive of memorizing the training data <http://bactra.org/notebooks/interpolation.html>, though if anything I did not talk up cross-validation _enough_.
HOISTED FROM OTHER PEOPLE'S ARCHIVES: Cosma Shalizi: The Rise of Intelligent Economies & the Work of the IMF: A reading from 2018-10-18: Still, I think, very wise on actually useful really-existing "AI" as very high-dimensional very big-data non-parametric regression & classification... <https://braddelong.substack.com/p/hoisted-from-other-peoples-archives>
Didn't some guy write an article titled "In Soviet Union, Optimization Problem Solves You"? There are two ways AI is being deployed.
In one approach, AI is used to help more advanced users work more efficiently and effectively. AI helps the more competent somewhat, but it helps the less competent more. This is not the stuff of which dream employees are made.
In the other approach, AI does the dumb, easy stuff, and humans stuck having to do the hard stuff like taking bolts from a box and putting them in an envelope. This is the future of AI as it will be applied.
Computers have enabled a level of command and monitoring beyond more traditional approaches, so we have truck drivers and warehouse workers with their movements determined by others and scheduled to the minute. AI is going to bring in a lot more optimization problem solves you. This will work well enough, as long as cheap, powerless employees are available.
There has been quite a lot of progress since 2017. ML can use training on one task and not start from scratch on another task, This was demonstrated by DeepMind on different videogames.
However, this doesn't detract from his main points which seem pretty relevant today. What was not anticipated then, is the hell-for-leather run to try to reach AGI as an AI goal. This has distorted AI funding and corporate goals, not to mention the blunt use of LLMs to displace labor with mostly inferior, but fast, boits. I gather agentic AI is the current focus (because profits?) rather than solving problems AI might be useful for. But capitalism...
Really existing AI is really cool, but implemented transformation takes much longer than expected. Perhaps there is so much capital flowing into so many firms that there will be no super-profitable hegemonic firm, particularly with China's inclusion. Three examples:
1) So many years since long-haul trucks would be driverless, even as distribution centers on Interstates in the exurbs have proliferated.
2) Waymo has been running for years in San Franciso, yet we wait on Tesla robotaxis to revolutionize automated driving.
3) So many LLM's available (OpenAI, Google, Meta, X, Claude, DeepSeek, Tencent, Baidu), along with private ones like Foxconn. Where's the moat? The value may be in application, not models.
Didn't Chile have some kind of up to the minute economic data collection system under Allende? It sounds pretty bogus, but a friend of mine ran into a guy involved in the implementation.
Cosma Shalizi emails:
> I appreciate the endorsement, and signal boost....
> 1. On Chile and Project Cybersyn: [http://bactra.org/notebooks/cybersyn.html]
> 2. On making people conform to machines, rather than the other way around: That's what I was trying to get at in the last paragraph. In a different context I'd have expanded on it, and been much more explicit.
> 3. On generalization ("transfer") across games: I think my remarks largely hold. There's been progress, but it's stuff like this <https://proceedings.neurips.cc/paper_files/paper/2022/hash/b2cac94f82928a85055987d9fd44753f-Abstract-Conference.html>, where they train on 41 related videogames and generalize to 6 other related videogames. The key point to my mind is that, to quote their own conclusions "Our results are based largely on performance in the Atari suite, where _action and observation spaces are aligned across different games_" (my emphasis). I will admit that I have not followed stuff on game-play as closely since then as I probably should be, so if Tolley (or anyone else) has pointers to papers, I'd be appreciative.
> 4. I was actually excessively dismissive of memorizing the training data <http://bactra.org/notebooks/interpolation.html>, though if anything I did not talk up cross-validation _enough_.
HOISTED FROM OTHER PEOPLE'S ARCHIVES: Cosma Shalizi: The Rise of Intelligent Economies & the Work of the IMF: A reading from 2018-10-18: Still, I think, very wise on actually useful really-existing "AI" as very high-dimensional very big-data non-parametric regression & classification... <https://braddelong.substack.com/p/hoisted-from-other-peoples-archives>
Didn't some guy write an article titled "In Soviet Union, Optimization Problem Solves You"? There are two ways AI is being deployed.
In one approach, AI is used to help more advanced users work more efficiently and effectively. AI helps the more competent somewhat, but it helps the less competent more. This is not the stuff of which dream employees are made.
In the other approach, AI does the dumb, easy stuff, and humans stuck having to do the hard stuff like taking bolts from a box and putting them in an envelope. This is the future of AI as it will be applied.
Computers have enabled a level of command and monitoring beyond more traditional approaches, so we have truck drivers and warehouse workers with their movements determined by others and scheduled to the minute. AI is going to bring in a lot more optimization problem solves you. This will work well enough, as long as cheap, powerless employees are available.
There has been quite a lot of progress since 2017. ML can use training on one task and not start from scratch on another task, This was demonstrated by DeepMind on different videogames.
However, this doesn't detract from his main points which seem pretty relevant today. What was not anticipated then, is the hell-for-leather run to try to reach AGI as an AI goal. This has distorted AI funding and corporate goals, not to mention the blunt use of LLMs to displace labor with mostly inferior, but fast, boits. I gather agentic AI is the current focus (because profits?) rather than solving problems AI might be useful for. But capitalism...
Really existing AI is really cool, but implemented transformation takes much longer than expected. Perhaps there is so much capital flowing into so many firms that there will be no super-profitable hegemonic firm, particularly with China's inclusion. Three examples:
1) So many years since long-haul trucks would be driverless, even as distribution centers on Interstates in the exurbs have proliferated.
2) Waymo has been running for years in San Franciso, yet we wait on Tesla robotaxis to revolutionize automated driving.
3) So many LLM's available (OpenAI, Google, Meta, X, Claude, DeepSeek, Tencent, Baidu), along with private ones like Foxconn. Where's the moat? The value may be in application, not models.
This is really good, thanks for hoisting it.
Didn't Chile have some kind of up to the minute economic data collection system under Allende? It sounds pretty bogus, but a friend of mine ran into a guy involved in the implementation.
See Dan Davies on Chile and Cybersyn
https://en.wikipedia.org/wiki/Project_Cybersyn
https://backofmind.substack.com/p/substacking-tolerances
Thanks. I had been wondering if this had been real, after a fashion.