Links #24
A flawed model of social media, new sleep gene results, and 3D printing innovation.
Not as much time to write this week so this post is light on explanation for each link.
1.
Daron Acemoglu, Asuman Ozdaglar, and James Siderius have a new paper AI and Social Media: A Political Economy Perspective. It formalizes everyone’s implicit model of echo chambers on social media. They assume people want to see content that they agree with, and the recommendation algorithm feeds them that while shaping their social network towards sycophancy. People update their views to be more extreme in response to seeing self-confirming evidence.
The problem is that a model like this isn’t borne out empirically. The video Fake News, Polarization, and Echo Chambers: What Science Says by Sabine Hossenfelder highlights research that very few people are stuck in echo chambers consuming media from only one political perspective.
But here’s the biggest problem with the model in the paper. Social media doesn’t get more engagement by showing you content you agree with but by showing you viewpoints you disagree with. Divisiveness drives engagement, not sycophancy.
The “polarization” section of the video above discusses some empirical evidence that exposure to other political view points increases polarization. It is known that facebook and other social media companies intentionally use controversy to drive engagement.
This video comes to similar conclusions and makes the argument that insulating ourselves from other political viewpoints would make the internet more pleasant and less polarized.
I can’t trust this paper because it’s missing basic facts about how social media works. I would like to see more economics and policy research on how to improve our information ecosystem that take into account these dynamics.
EDIT: see also Scapegoating the Algorithm—Asterisk
2.
Two new sleep results:
Multi-ancestry genetic architecture of sleep duration and its relationship to other sleep and psychiatric phenotypes | medRxiv. The total effect of the genes found in the study was 220 minutes of sleep or 3.5 hours. It seems that short sleep comes from having lots of different genes together rather than one short sleep gene.
Oveporexton, an Oral Orexin Receptor 2–Selective Agonist, in Narcolepsy Type 1 | New England Journal of Medicine. Promising results from one of Takeda’s narcolepsy trials.
3.
Novel High Resolution 3D Printing Method for Metals and Ceramics. 3D printing keeps improving. There’s lots of ways to potentially improve this idea, perhaps a dip in liquid nitrogen to freeze the slurry faster, or employing a solvent other than water so a faster supercritical drying step can be used. Freeze-pump-thaw steps and directional freezing could potentially reduce bubbles further. See also Rapidia’s metal paste 3D printing.
Everything else
Global Convergence of Gradient EM for Over-Parameterized Gaussian Mixtures. I haven’t read this yet but it matches a lot of keywords I’m interested in. The fact that overparameterized models have good convergence guarantees rhymes a lot with singular learning theory.
CLONE: Strap on a VR headset and teach a robot to move using your own body movements. We’re approaching scalable ways to produce much more data for robotic manipulation.
Mechanize, Inc is working towards automating labor with an emphasis on gathering quality data needed to complete individual tasks. This resonates with what I talked about at the end of AI Scaling.
Related: Hundreds of daily decisions illustrates just how much tacit knowledge there is in science. That makes it hard to automate. Standardizing science might be the path forward.
Thread by Auke Hoekstra speculating on how batteries and renewables can use digital synchronization techniques to make the grid more stable.
Another report showing that solar and batteries in sunny places can provide ~90% of electricity needs for competitive prices.
Bioprecipitation is neat. Some bacteria have surface proteins that nucleate ice crystals. These are a key cause of frost damage on crops. Genetically engineered ice-minus bacteria could compete with the wild type and prevent $1 billion of frost damage annually.
Keeping Ahead of Contagion. Nice review of the different technologies people are building to track outbreaks.
Ultra-wide-field, deep, adaptive two-photon microscopy. More progress on imaging brains. This method apparently can visualize neural activity in living specimens.
Mr. Secretary, Reclassify the Statin. Statins seem to be our best lifespan extension drug to date.
The fewer the merrier looks at the history of unified land ownership in Victorian era London. It points out that swathes of London that had a single owner built more public gardens and local public goods because those goods raise property values elsewhere. Some towns in the UK still have unified ownership. The author goes through some examples, discussing how each invests in local amenities to raise land values.
Why do commercial spaces sit vacant? How commercial property is financed and valued has some odd quirks and nuanced land tax policy needs to take this into account.
Multiverse Finance. Owning financial assets conditional on a certain future outcome. These markets are great for insuring against certain events. Though the additional complexity and risk has to be weighed against the benefits.
Matt Yglesias on debating. I agree, written debate is our best system for truth so far.
Child Tax Credits Pay for Themselves via Reduced Future Government Spending & Increased Tax Revenue
I got around to reading "Beggars in Spain" as recommended in your post on sleep. Thanks for the suggestion. It was an enjoyable read but it felt more like a meditation on editing intelligence than editing the need for sleep--the Sleepless/Sanctuary society was remarkable because of the intellectual & technological achievements of the members, not simply because they had roughly 1/3 more awake time.
Your couple of links on the competitiveness of solar + batteries makes me even more frustrated by the political outcomes of this weekend--not only gutting the investment tax credit but also adding in excise taxes on wind and solar. At least batteries are escaping relatively unscathed for now. It's a challenging week for energy, and I fear that by artificially curtailing our supply of cheap (and clean) energy we're effectively ceding the lead in AI to China. When electricity bills start creeping up in the coming years I can already imagine all of the headlines blaming data centers rather than the hamstringing of the clean energy development pipeline.