Commentary on CWT w/ Jack Clark - Bullish case for AI impact on economic growth

    Link to MR post: MR

    Jack and Tyler are both of the opinion that AI will cause lower economic growth than the 20%-30% estimates of those who are "feeling the agi" because real world bottlenecks in the world of atoms, which still makes up the majority of the Economy by far, will put some brakes on that breakneck acceleration.

    Now, I'm certain that both of them have already thought about the argument I'll make here, but I just want to spell it out loud. Many bottlenecks in the real world are consequences of political disputes (broadly speaking) in which parties have to rely on claims to support their side. In a world where more people rely on AI - and here I'm imagining the voting public using using AI to help them think about these disputes - it's possible that we'll see more rational, or at least better supported by evidence, policy getting an edge.

    For example, a supporter of a proposal to reduce the maximum speed in a road can now not only easily generate a report with proofs that such a measure on average reduce deaths where it's implemented, but they can also challenge anyone who opposes the validity of such a conclusion to try to extract an opposite one from their preferred LLM. Failure to do so should on average weaken less substantiated position (I'm not saying reducing speed limits are good, it's just an example and one could think of many better ones).

    Now, there are infinite reasons why that wouldn't be the case. I get that there are suspicions about bias in academia, actual bias in academia, people use motivated reasoning, people simply disregarding valid rationales for policies or outcomes that they don't like, etc. In short, I'm well aware that information being available is not a panacea for good policy. If it were, probably government would be a lot better than it is since it's not like there's a dearth of information in the Internet age.

    Going back to economic growth and AI, where I'm getting at is that AI availability should improve the ability for those arguing in good faith to convince genuinely undecided voters, which in turn should on average increase favor for actually well-founded policies, which in turn should translate to higher utility, which is very likely correlated with economic growth. I hope I'm adequately conveying how indirect I believe this effect would be. I certainly don't expect this effect to account for the 25% gap in economic growth estimates, but I believe it should improve somewhat on the base expectation that policy will drag things down by that much.

    The flip side, of course, is that AI might also make it easier for constituencies to vote *against* policies that might increase overall utility if it decreases *their* utility. It should become a lot easier to see through "trust me, this is going to be good for you" arguments that ultimately hurt them, but might increase GDP or something like that. I suppose we'll have to see the impact of both effects, but I believe in general people are aware of their own interest and care about others, albeit less than themselves (which I don't think is bad) and are genuinely interested in balancing these interests in a way that they think is fair, so I'll believe the net will be positive.

Comments

Post a Comment

Popular posts from this blog

JIT and GIL removal are not even my most anticipated Python 3.13 features

     Fun story: I came to Python through data analysis - so I wasn't a "python/programming expert" - and for an unreasonably long time used exclusively csv, parquet (thanks to pandas) and later JSON to persist data from my projects (local projects, of course. Mongo, RDS and Athena were always there). At work this was 100% enough (mostly parquet), but it always felt a little clunky to use JSON for a quick cache, especially when said cache grew larger for longer running processes or longer projects. Opening, parsing then serializing back again becomes slow surprisingly fast.      Then after a few years I decided to start using SQLite. I knew it existed, but always felt like it was a bit of a hassle to define a schema and "get" and "set" functions in all projects just to have a basic persistent k-v functionality. It was OK and had the extra benefit of allowing me to safely check this cache on a different jupyter notebook process when running longer proce...
Read more

An interview with Steve Wozniak by Jessica Livingston cured my AI anxiety

Note: there's no hidden knowledge contained in this post, it's just a reflection on how reading someone's reaction to an analogous situation changed my outlook Like everyone who works in an information job, I have been worried these last few months about how AI would impact my place in the labor force in the coming decade. Even though I didn't get to the depressed extremes we've been seeing on the Internet, I could not shake off the feeling that what I could offer to a company - no, to the whole economy, which just means "other people" - would get a lot less valuable. Come mid January, I wast taking a week off and had taken to read a few books since I read so little non-technical content on a day to day basis. One of these books was "Founders at work" by author and YC co-founder Jessica Livingston. The book is a series of interviews conducted by Jessica with the founders of some of the most important tech companies - sometimes specific products -...
Read more

What if regular exercise is the best cognitive exercise?

File this under highly speculative. I’m sure there is better research on this topic than this writing This past week, for no obvious reason, I considered a few thoughts at the same time: Exercise is supposedly one of the best anti-aging “drugs”, and also highly effective (relatively speaking) against cognitive decline There’s this vague idea that after you become “professionally inactive”, you accelerate your cognitive decline To avoid cognitive decline, one must keep “exercising the brain”, so the incentive for seniors to have hobbies, do crosswords, sudoku, etc Now, from this point it seems that I’d go to a very commonplace conclusion about how exercising is probably better than sudoku or something, and to be frank that’s exactly where I’m going. But another two things crossed my mind at the same time: Recent advances in AI, and less recent advances in computing in general showed that a lot of the things we consider “cognitively hard” - from symbolic manipulation to arithmetic, “...
Read more