Links for 2024-11-12
AI:
This study says it achieved human-level scores on the ARC challenge with LLMs using Test-Time Training (TTT). Can TTT combined with LLMs be enough to achieve abstract reasoning? https://ekinakyurek.github.io/papers/ttt.pdf [code: https://github.com/ekinakyurek/marc]
People underestimate how powerful test-time compute is: compute for longer, in parallel, or fork and branch arbitrarily—like cloning your mind 1,000 times and picking the best thoughts. https://www.tylerhouchin.com/blogs/entering-the-inference-era/
Improving semantic understanding in speech language models via brain-tuning — Speech models like Whisper can be improved by fine-tuning on brain data (fMRI collected by listening to podcasts). https://arxiv.org/abs/2410.09230
ManipGen: a sim2real agent for zero-shot manipulation. ManipGen handles complex tasks in the real world like organizing shelves, tidying cluttered tables, and more – all from text input and with on human demonstrations! https://mihdalal.github.io/manipgen/
Kinetix: an open-ended universe of physics-based tasks for RL https://kinetix-env.github.io/
Lucid V1:A world model that can emulate Minecraft environments in real-time on consumer hardware! https://ramimo.substack.com/p/lucid-v1-a-world-model-that-does
Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces https://arxiv.org/abs/2410.09918
Dreaming Out Loud: A Self-Synthesis Approach For Training Vision-Language Models With Developmentally Plausible Data https://arxiv.org/abs/2411.00828
Researchers say the path to wise AIs runs through metacognition https://arxiv.org/abs/2411.02478
Can AI models collectively match human forecasting abilities in predicting real-world events? The LLM ensemble outperformed the 50% baseline with a Brier score of 0.20 vs 0.25, matched human crowd accuracy (0.19 Brier score) with no statistically significant difference, showed 61% of predictions above the 50th percentile, and demonstrated significant forecast improvements of 17-28% when exposed to human median predictions, with prediction intervals narrowing from 17.75 to 14.22 for GPT-4 and 11.67 to 8.28 for Claude 2. https://www.science.org/doi/10.1126/sciadv.adp1528
Dario Amodei, CEO of Anthropic, in an interview with Lex Friedman: In the unlikely scenario where we proceed in a straight line without bottlenecks on data, compute, or energy, we could reach AGI in the next few years. 2026 or 2027; AI models are on a trajectory to surpass human reasoning and performance at professional tasks: "if we extrapolate the straight curve within a few years, we will get to these models being above the highest professional level of humans"; $100 billion AI data centers will be built by 2027 and he is bullish about powerful AI happening soon because we are starting to reach PhD-level intelligence; "the scaling is going to continue, there's some magic to it that we haven't explained on a theoretical basis yet" https://lexfridman.com/dario-amodei-transcript/
OpenAI's Kevin Weil says AI models "are going to get smarter at an accelerating rate" with voice and translation at a "magical" level already and vision to come https://youtu.be/IxkvVZua28k?si=0nJsYDyqDRCARPuk&t=2126
AI Takeoff Turns Data Centers Into America’s New Building Boom https://www.bloomberg.com/news/articles/2024-11-08/ai-takeoff-turns-data-centers-into-america-s-new-building-boom [no paywall: https://archive.is/DiFjM]
Amazon steps up effort to build AI chips that can rival Nvidia. https://www.ft.com/content/3d9b5c6d-f1ae-4f6f-adc3-51e5f1dfb008 [no paywall: https://archive.is/DLDmJ]
“The FrontierMath benchmark does something different from the IMO and Putnam” https://blog.evanchen.cc/2024/11/10/frontiermath/
What if AI could book your appointments, collaborate with others, and navigate life’s complexities on your behalf? The Ethics of AI Assistants with Iason Gabriel https://www.youtube.com/watch?v=aaZc-as-soA
Bayes3D: fast learning and inference in structured generative models of 3D objects and scenes https://arxiv.org/abs/2312.08715
AI-driven mobile robots cooperate to conduct chemical synthesis https://www.theengineer.co.uk/content/news/ai-driven-mobile-robots-cooperate-to-conduct-chemical-synthesis
“I Went Birding With the World’s First AI-Powered Binoculars” https://www.wired.com/story/swarovski-optik-ax-visio-ai-binoculars/ [no paywall: https://archive.is/EeToa]
How ChatGPT Brought Down an Online Education Giant: Chegg’s stock is down 99%, and students looking for homework help are defecting to ChatGPT https://www.wsj.com/tech/ai/how-chatgpt-brought-down-an-online-education-giant-200b4ff2 [no paywall: https://archive.is/3nUKF]
YUDKOWSKY + WOLFRAM ON AI RISK. https://www.youtube.com/watch?v=xjH2B_sE_RQ
Prediction markets:
Google difficulties in forecasting LLMs using a internal prediction market https://asteriskmag.com/issues/08/the-death-and-life-of-prediction-markets-at-google
The Online Sports Gambling Experiment Has Failed https://www.lesswrong.com/posts/tHiB8jLocbPLagYDZ/the-online-sports-gambling-experiment-has-failed
Health:
"James Fickel has dedicated $200 million he made betting on Ether to becoming one of the world’s biggest investors in" longevity & neuroscience https://www.bloomberg.com/news/articles/2024-11-11/crypto-millionaire-fuels-push-to-transform-brain-research [no paywall: https://archive.is/VhxMU]
When muscles work out, they help neurons to grow, a new study shows https://news.mit.edu/2024/when-muscles-work-out-they-help-neurons-grow-1112
What Ketamine Therapy Is Like https://www.lesswrong.com/posts/zgAws2AoFE3adigvy/what-ketamine-therapy-is-like
Technology:
"Neural Networks (MNIST inference) on the “3-cent” Microcontroller" (90% MNIST in 1 kiloword) https://cpldcpu.wordpress.com/2024/05/02/machine-learning-mnist-inference-on-the-3-cent-microcontroller/
Shocking New Memory Tech: Crystal-to-Glass Transformation Using a Billion Times Less Energy https://iisc.ac.in/events/self-shocks-turn-crystal-to-glass-at-ultralow-power-density/
Space:
“The radio source ASKAP J1935+2148 is an amazing thing…It could be a really weird pulsar, but nobody knows how a pulsar spinning so slowly could put out radio waves.” https://mathstodon.xyz/@johncarlosbaez/113460684713479446
“Why Sabine Hossenfelder is Just Wrong” https://www.math.columbia.edu/~woit/wordpress/?p=14232
Ukraine:
Butter thefts highlight cost of Russia’s war economy. Rampant inflation of staple food prices is linked to soaring defence spending. https://www.ft.com/content/659bb41c-d6f1-4690-8193-647f549d5133 [no paywall: https://archive.is/iJNgr]
DeepState about the massive Russian attacks on the Kursk front. Up to 300 Russians were killed or injured https://x.com/bayraktar_1love/status/1856111073159496113
“Very bloody battles are going on in Kursk oblast. After a failed attack at the first day of operation the enemy used a bare minimum in terms of using AFV at my flank and in the centre.” https://x.com/OSINTua/status/1856030625880588444
