Links for 2022-11-29
Write-up on deciding which DNA testing service you can purchase, based on what you're hoping to learn by Razib Khan. https://razib.substack.com/p/so-you-want-to-test-your-dna
"Can Foundation Models Talk Causality?", Willig et al 2022 (causal explanations generated by GPT-3, OPT, & AlephAlpha’s Luminous) https://arxiv.org/abs/2206.10591
Solving math word problems with process- and outcome-based feedback: “We improve accuracy over raw chain-of-thought from 41% to just 13% error.” https://arxiv.org/abs/2211.14275
“MagicVideo can generate video clips with 256×256 spatial resolution on a single GPU card, which is 64 × faster than the VDM” https://magicvideo.github.io/
SinFusion: Training Diffusion Models on a Single Image or Video https://yanivnik.github.io/sinfusion/
“US government estimates that the level of human activity in cislunar space over the next decade could exceed everything that’s happened there between 1957 and today, combined.” https://vox.com/platform/amp/recode/2022/11/22/23473483/white-house-joe-biden-moon-artemis-permanent-outpost-spacex
"By inserting a gene into the cells of the optic nerve, we can make it possible to stimulate them using a small display inserted into the eye." https://science.xyz/news/announcing-the-science-eye/
“asteriskmag.com, the new rationalist / effective altruist magazine.” https://astralcodexten.substack.com/p/is-wine-fake-in-asterisk-magazine
“William MacAskill’s latest book presents itself as an introduction to the burgeoning longtermist movement. But his views are eccentric – even within the movement he founded.” https://asteriskmag.com/issues/1/review-what-we-owe-the-future
"Contrary to the deterioration hypothesis, we find that market-oriented societies have a greater aversion to unethical behavior, higher levels of trust, and are not significantly associated with lower levels of morality." https://www.sciencedirect.com/science/article/abs/pii/S0167268122003596
“Can we learn Deep Neural Networks that explain their own predictions using human-interpretable concepts? Yes!” https://arxiv.org/abs/2205.15612
Preventing an AI-related catastrophe https://80000hours.org/problem-profiles/artificial-intelligence/
By providing algorithm details in-prompt, rather than just a couple examples, LLMs can actually do addition up to 19 digits!
Even when prompt only includes examples with < 5 digits.
Paper: https://arxiv.org/abs/2211.09066
Elbit's micro-suicide drone swarms can hunt enemies in urban combat:
If it wasn't for environmentalists, we could have already defeated climate change.
Just imagine if instead of pumping billions into renewables, this money would have gone into researching next-generation nuclear power. Imagine if nuclear power would have been deregulated instead of massively overregulated. It would be orders of magnitude cheaper and it would take fewer than 5 years to build a new nuclear power plant.