Links for 2023-02-24
A “miracle antibiotic” in the offing? New compound works against all multiple-drug-resistant bacterial strains, and also stymies the evolution of bacterial resistance https://whyevolutionistrue.com/2023/02/22/a-miracle-antibiotic-in-the-offing-new-compound-works-against-all-multiple-drug-resistant-bacterial-strains-and-also-stymies-the-evolution-of-bacterial-resistance/
“Scaling Robot Learning with Semantically Imagined Experience: ...we show that manipulation policies trained on data augmented this way are able to solve completely unseen tasks with new objects and can behave more robustly w.r.t. novel distractors.” https://diffusion-rosie.github.io/
Wonder why you can’t put down your phone? Reinforcement Learning for User Retention (RLUR) might be to blame: “Techniques like RLUR are societal change in an acronym trenchcoat; this is how we build systems to automatically harvest the attention of people across the world – not with a bang, but with backpropagation!”— Jack Clark https://arxiv.org/abs/2302.01724
RealFusion 360◦ Reconstruction of Any Object from a Single Image https://arxiv.org/abs/2302.10663
“Travelling from the surface of Earth to Earth orbit is one of the most energy intensive steps of going anywhere else. This first step, about 400 kilometers away from Earth, requires half of the total energy needed to go to the surface of Mars.” https://www.nasa.gov/mission_pages/station/expeditions/expedition30/tryanny.html
Andrej Bauer on all the “invisible” parts of a mathematical statement that human mathematicians implicitly fill in when communicating with each other, but which formal proof assistants can struggle with. https://www.youtube.com/watch?v=wZSvuCJBaFU
Hippocampal spatial representations exhibit a hyperbolic geometry that expands with experience https://www.nature.com/articles/s41593-022-01212-4
How much space would we need to store the world’s plastic in landfills? https://hannahritchie.substack.com/p/landfill-plastic-area
“The occurrence of Zugunruhe even in cage-raised birds with no environmental cues (e.g. shortening of day and falling temperature) has pointed to the role of circannual endogenous programs in controlling bird migrations.” https://en.m.wikipedia.org/wiki/Bird_migration
“The Ishango bone, discovered at the "Fisherman Settlement" of Ishango in the Democratic Republic of Congo, is a bone tool and possible mathematical device that dates to the Upper Paleolithic era.” https://en.m.wikipedia.org/wiki/Ishango_bone
“Dogs looking similar to Salukis are shown on wall carvings of the Sumerian empire (now Iraq), dating from 6,000 to 7,000 BC. The ancient skeletal remains of a dog identified as being of the greyhound/saluki form was excavated at Tell Brak in modern Syria, and dated to approximately 4,000 years before present.” https://en.m.wikipedia.org/wiki/Saluki
Fair and Square: A Conclusion on IQ Test Bias https://menghu.substack.com/p/fair-and-square-a-conclusion-on-iq-test-bias
The genie is out of the bottle:
1. LLMs are already powerful enough to solve a lot of real-world problems when you chain them together and grant them access to interpreters and other tools.
2. The remaining challenges towards weakly general intelligence can and will be overcome relatively soon.
3. It is too late to censor this technology because the most important puzzle pieces are already out there in the wild.
P.S. Agency will emerge naturally:
1. https://gwern.net/Tool-AI (The bottleneck for AI is humans. Humans in the loop make it slow. Economic incentives will want to remove humans and give AI increasing autonomy.)
2. https://pastebin.com/U7quKPhu (Agency is part of the world. If you want to predict the world, you need to model agency.)
In 2010, people were still debating whether an AI could break out of air-gap confinement by convincing its creators via a text interface to let it out of the box. Everyone expected that the researchers would at least be careful enough to prevent their AI from interacting directly with the world.
2023: “Now that AI can read your code on Replit, your errors, and can even eval in the REPL it can send you fix diffs”
Yudkowsky: "Look at Sydney's suggested "user responses" at the bottom, after the main chat tries to apparently terminate Sydney's side of the conversation."
User @the__unraveling explains: "he's referring to Bing's apparent attempt to override the safety mechanism being triggered and continue the conversation via the suggestions it makes at the bottom of the screen"