MAY-jor developments

The best content that May had to offer

1. Change My Mind: Density Increases Local But Decreases Global Prices
   Scott Alexander with a contrarian take that increasing the supply of housing in the cities people most want to live will increase prices within those cities due to induced demand effects. On the one hand, I think he is right to some degree. There is so much pent-up demand for housing in “big” cities that increasing the supply of housing by an additional 20,000-30,000 would probably not make any difference, and because everything is rarely equal, prices may actually increase.
   On the other hand, as the economist Scott Sumner, replies, even if Scott’s argument is true the solution remains the same: build more housing. Increasing supply will still have aggregate effects on prices even if the local effect goes in the opposite direction. Furthermore, the goal is to increase supply in all cities, not just the big cities.

2. AI experts aren’t always right about AI [Paywalled]
   Tyler Cowen maintaining his anti-doomer position, which I believe is justified. After reading him several times on this topic, I think I can summarize his position as follows:
   a) Every time a new innovation pops up people fear the worst, even very smart people. They have always been wrong and superintelligent AGI does not appear likely to be the exception.
   b) Intelligence is not the all-powerful thing AI experts believe it is.
   c) No one is considering the tradeoffs. If there is some chance it brings the end of humanity, but a much larger chance it generates unprecedented levels of human flourishing to googolplexes of future humans, shouldn’t we take that bet?
   d) Asking gerontocracy governments who can barely work their Iphones to regulate AI is the worst of all possible outcomes. It will not only fail, but make everyone much worse off in the process.
   While I somewhat disagree with a), and am unsure about b), the other two are hard to disagree with (though I expect his priors on c) are much different than skeptics here). It does seem that inventing something that could potentially be orders of magnitude smarter than any human in every way is vastly different from nuclear weapons or the printing press.
   As for whether AI experts are overrating intelligence, I struggle to appreciate the implications of their arguments, but also wonder whether I’m just failing to think big enough. If what they are saying is true, the only thing preventing Von Neumann from being able to solve every problem and being able to rule the world was a bit more intelligence. Obviously, the comparison isn’t fair: machines don’t need sleep, they don’t have emotions, and can edit their own source code to recursively self-improve. Still, I really do doubt that any superintelligence can rule the world, and I suspect that anyone who says otherwise might be underestimating the complexity of the world.

3. Our Alien Stalkers
   Robin Hanson exploring the possible existence of extraterrestrial life through his grabby aliens model continues to be one of my favourite things, even if it turns out to be pointless.

4. Your IQ isn’t 160. No one’s is.
   Has some good points, but I think a lot of it is based on misconceptions and an incorrect interpretation of the data. The main misconception is that IQ and intelligence are the same thing- they’re not. We cannot observe or measure intelligence directly, but we can use IQ tests to approximate it. So yes, you may have someone that scores 160 on an IQ test that is generally not that smart, but just because someone got lucky doesn’t undermine the importance of intelligence, nor the general validity of intelligence tests. Thus, IQ score might be changeable through practicing for IQ tests, but intelligence doesn’t seem to be.
   Also, he uses Einstein’s grades as an example of how intelligence tests don’t matter. This only works if you believe intelligence = being good at taking tests. Einstein was obviously extremely clever, and his early grades appear more likely to be a symptom of boredom rather than any lack of ability.

5. What Midjourney thinks professors look like, based on their department
   Thank me later.

6. We Have No Moat, And Neither does OpenAI
   TL;DR a leaked an internal memo from Google claiming that big companies like Google and Open AI can’t compete with open source, since it can innovate faster without the being constrained by the bureaucratic bloat that comes with being a part of a big company.
   Whilst I disagree with the author’s main point because there are obvious advantages to having more compute, it did help me appreciate the significance of LLaMA getting leaked and LoRA.

7. Why is spending so resilient?
   This post questions why the economy is seemingly doing okay despite almost everything going up in flames, and makes the case for a limit cycle view. In the author’s own words:
   ”the limit cycle view suggests that a powerful enough shock, such as the rebound when economies re-opened after the pandemic recession, can put us on a upward spiral, and that such spiral features such a strong self-reinforcing mechanisms that even large negative shocks cannot derail us”.
   I’ve heard a few people refer to macroeconomics as more art than science, and reading this post certainly nudged me further in that direction.

8. Top AI Tools for Business and Work
   Some of the tools in here are outstanding, especially some of the memory ones which tracks every action you make on your device all day which you can then ask questions about. Numerous.ai is also good for making low-cognition-but-time-consuming Excel tasks a lot quicker, and a little insight into what Microsoft Co-Pilot might look like. The sad thing is, I can use none of these tools for work, despite the fact that they would make me at least 2x more productive.

9. Unlimiformer: Long-Range Transformers with Unlimited Length Input
   Another potentially big development in the AI space that shatters the token limit. If my understanding is correct, instead of processing each token to help predict the next one (causing the computational cost of each additional token to be n^2), they use a K-nearest neighbours algorithm on the Transformer’s encoder output to find the most relevant k tokens, thus making the cost of each token increase sublinearly. Probably pretty useful for writers who want LLMs to help edit long books.

10. AI Alignment as a Solvable Problem | Leopold Aschenbrenner & Richard Hanania
    Probably the most sane takes I’ve seen on the X-risks of AI. Leopold views things in much the same way I do: alignment is an engineering problem, not a philosophy one. Pontificating different x-risk scenarios on LessWrong that are based on a series of tenuous assumptions mostly doesn’t get us anywhere.

11. Miles Turpin Twitter thread on chain-of-thought explanations from LLMs
    Not to anthropomorphize LLMs, but it’s interesting that they appear to reason in much the same way humans do: use their intuition to make their minds up about something, and then conjure up some reasoning to justify their feeling. Not sure if this is says more about the difficulty of alignment, or the difficulty of knowing thyself.

12. Killers of the Flower Moon trailer
    Another Scorcese and DiCaprio combo that’s sure to be a hit. Not merely because they are both great at their craft, but because the source material is incredibly good.

13. Alignment and Competition - with Robin Hanson and Agnes Callard
    This is the first time I think I finally understood that Robin’s critique of AGI doom is a bit complicated. He’s arguing against FOOM for inductivist reasons, but for death in the long-run due to the incomprehensiveness of alignment. What would it mean for a machine to aligned to “human values”? Well, first you would need everyone to agree what those are, and then give the machine a loss function that optimizes for those values. Except, people disagree over their values all the time, so inevitably one set of values would win out at the expense of everyone else’s, probably leading to doom for those with the “wrong” values (For instance, if the CCP get AGI first solve for the equilibrium).
    But let’s say you find a way to solve that problem, you are still screwed in the end since values are dynamic. The values of most people now are different than they were a century ago, and values can change dramatically in a short space of time. Consequently, even succeeding at aligning a superintelligence to human values today gives no guarantee that will mean anything tomorrow.

14. Why Britain doesn’t build

    Great history of the roots of the UK’s NIMBYism. The one thing that amazes me is the continuity between now and then: it is still the same kinds of people that will do whatever they can to stop poor people living anywhere near them. The sheer volume and proneness to activism of these people means that any democratic system is pretty f***ed until the NIMBY’s pass away from old age.

15. Language Models Meet World Models: Embodied Experiences Enhance Language Models
    Not sure if this is significant or not, but I suspect it could be.

16. College quality as revealed by willingness-to-pay for college graduates
    This is mostly confirmation bias on my part, but it has been my intuition ever since I started applying to graduate jobs. Once you control for how smart the student is, differences in eventual outcomes/productivity/salary mostly disappear. That’s why I think as a society we should think twice before we encourage all of our kids to go to university, because even if they graduate with a good enough grade, from a great university, the still need to be cognitively capable enough to handle the kinds of intelligence tests most employers put out (most of which are surprisingly easy, I might add).

17. Diminishing Returns in Machine Learning Part 1
    Love Brian’s contrarianism, and whilst I don’t buy the main thrust of his argument that AGI will be bottlenecked by compute (I think it will be bottlenecked by lots of other things, but not that), his timelines for superintelligent AGI are pretty much the same as mine. One thing he mentions that I don’t think people are factoring in is the regulatory ragnarok from an army of job-insecure midwits that is on the way.

18. Sibling variation in polygenic traits and DNA recombination mapping with UK Biobank and IVF family data
    Ever wondered why one of your siblings seems just like you? Or, on the contrary, why one of them is so different to you despite having the same parents? This paper answers: the degree of genetic variation between siblings is normally distributed. It is typically accepted as dogma that you share 50% of your DNA with each parent and your siblings, but the relationship apparently looks much more like this:

    

    Turns out your kid may just be very different from you due to random genetic variation. Depending on who you are, that may not be a bad thing.

19. Is Porn Misleading Men?
    For those not in the know, Aella is a former sex-worker who now performs provocative research for fun and posts it on Substack. The results of her survey here, rather counterintuitively, suggest that porn makes men more accurate in their predictions of what women want in the bedroom.
    On the one hand, I think a lot of the more surprising graphs in the post are the product of the idiosyncrasies of her sample of Twitter followers. On the other hand, there does seem to be signal here, which is further confirmed when she just looks at the people she paid to do the survey. Weird.

20. Eve Theory of Consciousness (v2)
    Fascinating theory on why and how consciousness evolved. Never did I ever think Julian Jaynes and the Bicameral Mind would make something of a comeback, but here we are. To be clear, I think is origin story is unlikely to be completely accurate, but the idea about consciousness starting in women appears highly plausible.