Links #23
And an update on the funding round for sleep need therapies.
1.
At the end of my post on sleep need reduction I mentioned a funding proposal for a self-experiment. It’s now up! We’ve managed to raise enough to do an initial trial of intranasal Orexin-A on acute sleep deprivation. I’m very excited to get started on this.
2.
Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? Blog explainer here. Repeatedly sampling the base model many times often eventually gives you the correct answer. Base models start to outperform RL approaches past a certain number of samples. The key here is to be able to sift through the samples to select the right answer.
In highly verifiable domains it’s easy to identify a correct answer. This is ironic because these verifiable domains were the ones where self-play and synthetic data looked most promising. Now RL is on a knife-edge: if the domain is highly verifiable you can just sample more to find the right answer, if it’s not, how are you going to generate training data?
RL is still exciting for other reasons, it seems to work well in practice, it’s easier to correct and audit, and it enables us to use smaller models.
3.
I think heatwaves are one of the biggest near term issues we’ll see from climate change. One of the big downsides of heatwaves is their impact on agricultural productivity. I’ve previously touted solar radiation management (SRM) as an option to delay the effects of climate change, so would SRM improve crop yields? There’s a lot of ways this could go:
Heatwaves are bad for crops, so keeping the planet cool should increase productivity.
But a hotter climate means more rain in most places, so cooling the climate can hurt rainfed agriculture relative to a climate change baseline.
SRM reduces sunlight reaching the surface, which may harm photosynthesis.
But plants already have to carefully control the amount of light they absorb to prevent damage from too much sunlight.
Carbon Brief covers one paper that looks at the actual effects of volcanic eruptions (which, among other things, inject SO2 into the atmosphere) and uses that to inform their climate model, finding that SRM would leave yields basically the same as if we just let climate change happen. On the other side of the ledger, these papers find small positive effects on yields:
Climate and carbon cycle response to the 1815 Tambora volcanic eruption
Solar geoengineering can alleviate climate change pressures on crop yields
The last of these is interesting because it finds that emissions reductions reduce yields because of lower CO2 fertilization.
These three papers gesture at the idea of a Goldilocks zone where a little SRM is good, but more is harmful to yields:
Current benefits of wildfire smoke for yields in the US Midwest may dissipate by 2050
Atmospheric opacity has a nonlinear effect on global crop yields
Optimal climate intervention scenarios for crop production vary by nation
Also note that crop productivity can be a proxy for the success of the wild plants upon which everything else depends.
Overall I lean towards the conclusion that a low level of SRM would increase yields. But it sure would be nice to have more research on this.
Related: Potential effects of climate change and solar radiation modification on renewable energy resources. Could SRM hurt solar by blocking light? Or make it better by keeping panels cooler? EDIT: this post by Austin Vernon suggests SRM won’t be a problem for solar generation and could even improve it in some ways.
Everything else
Power Lies Trembling synthesizes 3 different books: starting with a model of military coups as a social coordination game, applying this lesson to social preference cascades, and framing Kierkegaard’s knight of faith as someone who understands these principles and commits to their own beliefs.
Learned pain as a leading cause of chronic pain. I think this is a really important area of research. To what degree can chronic pain be addressed with (say) cognitive behavioral therapy?
Supernuclear is writing about how to live near friends and build local communities that make you happy.
Nice calculator on offsetting the impact of your meat consumption with charitable donations. I haven’t looked into how they do the calculations so be skeptical.
Lepodisiran — A Long-Duration Small Interfering RNA Targeting Lipoprotein(a). This is exciting! Heart disease kills more people than cancer, but it feels understudied. This si-RNA treatment provides long-acting, effective treatment for the ~10% of people who have genes that overproduce lipoprotein(a). Interesting that they were able to modify it to last a long time. RNA usually gets digested pretty quickly.
Autoimmune therapies: an inspiration for anti-aging therapies? Designer immune cells used in CAR-T therapy are inspiring new auto-immune disease therapies. Speculates that this could be used for anti-aging therapies as well. Designing cells to reprogram your immune system or destroy senecent cells sure sounds like nanotech to me.
Hyperspectral reporters for long-distance and wide-area detection of gene expression in living bacteria. Can this be used for live in vivo brain recording?
Whole-body physics simulation of fruit fly locomotion. Didn’t the whole connectome of a fruit fly get published recently? How close are we to uploading fruit flies?
Robot Dexterity Still Seems Hard. One big obstacles to deploying robots in the real world.
Paper2Code: Automating Code Generation from Scientific Papers in Machine Learning. Remember, ML took off because it was easy to copy and tweak other’s papers. This accelerates that.
Nonproliferation is the wrong approach to AI misuse. Trying to ban AI as a way to stop AI misuse is a bad idea because it curtails freedom and stifles innovation. Instead, we should develop mitigation technologies (perhaps with the help of AI itself). Delaying the proliferation of AI by a small amount (adaptation buffers) can help deploy these mitigations.
Space-based missile defense is a cool idea that never really panned out. With falling launch costs, could it make sense to develop now? No, argues this post from Naval Gazing “ space-based missile defense … is a very poor way to provide protection against nuclear attack to the United States. We should instead focus on existing ground-based systems as a faster and surer way to protect against missile attack.”
The Unsustainability of Moore’s Law lists obstacles to scaling chip fabs further, it doesn't look pretty (though industry has 10-20 years of innovation left). Author speculates that the paradigm might shift to simpler fabs and chips getting reused like cars instead of updating all the time.
Active energy compression of a laser-plasma electron beam. Plasma wakefields give us the opportunity to accelerate electrons and ions to high energies without huge equipment. The problem so far is that the energy spread of the resulting particles makes them less useful for research. This paper employs several techniques to bring that spread down. Could these beams of electrons be used as a higher throughput form of electron beam lithography?
This Physicist Says Black Holes are Quantum Computers. Paper here.
Nice thread on how rollback netcode works to keep videogames synchronized.
Is Dollar Dominance Good for the U.S.? Good post about the pros and cons of dollar dominance and how that interacts with national debt and trade deficits.
# The Neoliberal Climate Solution Nobody Wants to Hear
The best way to manage climate impacts is deregulation and free trade—essentially, a second neoliberal revolution. Moving to larger farms with better technology like robotic pest control and GMOs will enable more yield and better ability to adapt to changing climates. Free trade will ensure that even if crop failure happens in one country, it gets mostly nullified by imports from a neighbor, because drought and other extreme weather events tend to be uncorrelated.
Unfortunately, I'm pessimistic about these sets of reforms panning out. You know what Saudi Arabia, Venezuela, most of Europe, and India have in common? Complete GMO bans. This probably makes hippie environmentalism the most successful religion in human history.
The only countries that have GMOs are America (for now), China (of course), Australia (as God intended), and surprisingly Bangladesh—we wanted cheaper eggplants, okay?
If you read about the Asian rice inflation crisis of the late 2000s, it makes one very pessimistic about the ability of nations to cooperate on trade to mitigate climate challenges. These countries made the problem dramatically worse through protectionist responses that should have taught us better lessons about the importance of maintaining open agricultural markets during times of stress.