Links for 2022-10-19
“Over at Google, large language models have been plugged into physics simulators to help them share a world model with their human interlocutors, resulting in big performance gains. They call it Mind's Eye.” https://www.lesswrong.com/posts/SbadvzWbufzX9iWJf/they-gave-llms-access-to-physics-simulators
gwern on the China (and therefore Taiwan) semiconductor situation: “…if the CCP tries to invade Taiwan in the next 10 years, historians are likely going to point to 2-3 days ago as the pivot: it is now a razor blade cutting China's throat if AI and high tech in general is the future, with nothing more to lose and everything to gain from destroying TSMC so no one else can have it either.” https://www.lesswrong.com/posts/oBTkthd7h8sDpkiu2/analysis-us-restricts-gpu-sales-to-china?commentId=xS8fc82cLd83y6pbb [This is also related to the concept of an anthropic shadow: if artificial intelligence was to cause human extinction but required a lot of computing power, you would be more likely to find yourself in world lines in which the necessary conditions for cheap compute are not met. If you were living in such a world line you might expect to see crypto miners causing a GPU shortage or supply chain disruptions due to a pandemic. A war between the United States and China over Taiwan in which important chip fabrication plants are destroyed is also much more likely to occur in world lines that are not wiped out. An anthropic shadow hides evidence in favor of catastrophic and existential risks by making observations more likely in worlds where such risks did not materialize, causing an underestimation of actual risk.]
“It's been another fascinating ~week in twitter biotech. Going back through some highlights...we're stepping towards being able to generate arbitrary geometries of either synthetic (drugs) or organic (protein) parts” https://threadreaderapp.com/thread/1580748480037154817.html
“How Trustworthy Are Supplements?” “Pretty trustworthy, actually” https://astralcodexten.substack.com/p/how-trustworthy-are-supplements
‘Monumental’ Math Proof Solves Triple Bubble Problem and More https://www.quantamagazine.org/monumental-math-proof-solves-triple-bubble-problem-and-more-20221006/
"ReAct: Synergizing Reasoning and Acting in Language Models", Yao et al 2022 (PaLM-540B inner-monologue for accessing live Internet APIs to reason over, beating RL agents) [via @gwern] https://arxiv.org/abs/2210.03629
Understanding HTML with Large Language Models https://sites.google.com/view/llm4html/home
New gene editing technologies from Arc. (Better than CRISPR for some applications.) https://arcinstitute.org/blog/genome_insertions
“People working on suffering risks or s-risks attempt to reduce the risk of something causing vastly more suffering than has existed on Earth so far.” https://80000hours.org/problem-profiles/s-risks/
What Every C Programmer Should Know About Undefined Behavior #1/3 http://blog.llvm.org/2011/05/what-every-c-programmer-should-know.html
"the more costly it is to dispute X the less confident we can rationally be about X" https://philpapers.org/archive/JOSTES-2.pdf
Examples of the unreasonable usefulness of pure mathematics:
1. Logic was considered a hopelessly abstract subject with no conceivable applications until Claude Shannon turned it into a trillion-dollar business that is now underlying most of our technological civilization.
2. Pythagoras could not have imagined the uses to which his equation would be put.
3. The people who formalized complex analysis did not predict its usefulness in physics, including the branches of hydrodynamics, thermodynamics, and especially quantum mechanics.
4. Number theory was considered useless until its importance for cryptography was discovered.
5. Conic sections, developed in the 2-nd century BC, had no applications until Kepler's discovery that celestial bodies move on conic sections.
6. The history of fractals traces a path from chiefly theoretical studies to computer graphics and fractal analysis.
7. The Radon transform, when introduced by Johann Radon in 1917, was useless until Cormack and Hounsfield developed Tomography in the 60s (Nobel prize for Medicine 1979).
8. Many mathematicians regarded negative and complex numbers as absurd and useless before the 15th century. For instance, Chuquet referred to negative numbers as "absurd numbers." Michael Stifel has a chapter on negative numbers in his book "Arithmetica integra" titled "numeri absurdi". And so too were complex/imaginary numbers. Gerolamo Cardano in his book "Ars Magna" calls the square root of negative numbers a completely useless object.