Action Potentials for October

Note: I haven't done deep dives on all of these links. Corrections, criticisms, castigations, beratement, &c, are welcome. 

1: A team mostly from the Stanford psychiatry department has transplanted intact organoids into the developing cerebral cortex of rats. These integrated functionally into the rat brain and affected behavioral responses. Several of the authors declared a provisional patent application, making it seem like this research is potentially leading towards the use of organoid transplantation in neurosurgery. It probably would be much harder to get integration if the brain were not developing, but that’s an open question and potentially a fixable problem, and anyway plenty of neurobiological disorders are developmental.

Figure 1 s41586-022-05277-w

2: Dynomight does calculations on the expected lifespan loss of a single CT scan, and finds it is usually on the order of minutes to hours, making it not something to worry about most of the time. The estimates are noisy, though.

https://dynomight.net/scans/

3: An argument by Lawrence Newport that effective altruists should draw insights from the historical successes of Quakers, i.e. in accepting criticism, inviting debate, and trying to make change as a minority group in society.

4: RCT in children 7 to 15 years old (n = 99), most of whom had anxiety disorders, finds that a combination of CBT + sertraline was no better than CBT + placebo at remission or improving anxiety symptoms. The remission rate for the placebo pill group was high, 69% at a 6-month follow-up. They argue that a previous similar study did not adequately control for expectancy effects.

5: Ellen Pasternack on parenting as a public good.

6: A study finds that a model with 12,111 SNPs accounts for nearly all of the common SNP-based heritability for height in people of European ancestry. Forgoing a rare opportunity to say “no more research is needed”, which pretty much only Cochrane reviews will ever say, they point out more research is needed to reach equivalent saturation in other ancestries.

7: Cosma Shalizi on why you should quit Twitter. Argues that all online communities reproduce the problematic social dynamics of print literary communities and that Twitter raises this to a new level of awfulness.

8: Argument by Sam Atis that education actually seems helpful when you use regression discontinuity designs rather than trying to adjust for confounders, which we all know basically doesn’t work. Maybe education is not just signaling after all.

9: A welcome change in state law at the end of last year, which I just learned about, makes it easier to access medical aid in dying in California by decreasing waiting periods.

10: Resources for the Highly Sensitive Person.

11: In many places, taxes on alcohol haven’t even come close to keeping up with inflation, largely because of industry pressure. To wit, the cheapest liquor in 2010 cost 1/15th of what it did in 1950. This is even though we know that higher alcohol taxes are associated with lower rates of alcohol-associated disease, injury, and death. Why do we tax cigarettes and cannabis so much but have such limited taxes on alcohol, which arguably has the worst externalities? Anyway, taxes are necessary, so I’m always on the look out for areas where taxes should be higher, and this seems like an obvious place that would also reduce suffering.

12: Deepvail is a company founded in 2019 that does independent assurance of AI technologies in high-risk settings such as healthcare. My guess is that this is going to become a massive industry in the coming years as deep learning, or its successors — the key element being low interpretability AI methods — continues to eat the world, but people want to know whether it is safe before they deploy it.

13: Frequently seen but rarely diagnosed: musical ear syndrome (2016). Seems to be like Charles Bonnet syndrome (visual hallucinations in visually impaired patients), but for sounds.

14: The Eighty-Five percent rule (2019): some evidence suggests that learning may be best when accuracy on a task is not too hard but not too easy. A useful ballpark number of the optimal accuracy seems to be at around 85%. In my opinion, this has face validity, at least for human learning.