Links for 2024-11-06

AI:

1. Through a Glass Darkly: Mechanistic Interpretability as the Bridge to End-to-End Biology — How the single-cell field is on the precipice of its own AlphaFold moment https://www.markov.bio/research/mech-interp-path-to-e2e-biology

2. WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning — It improves the success rate of Llama-3.1-8B from 4.8% to 42.4%, and from 6.1% to 43% for GLM4-9B. The open models significantly surpass the performance of GPT-4-Turbo (17.6%) and GPT-4o (13.9%). https://arxiv.org/abs/2411.02337

3. Eurekaverse 🌎: A path toward training robots in infinite simulated worlds! Eurekaverse is a framework for automatic environment and curriculum design using LLMs. This iterative method creates useful environments designed to progressively challenge the policy during training, enabling the learning of complex skills. The automatic curriculum designed by Eurekaverse enables gradual learning of complex parkour skills in simulation and can successfully transfer to the real-world, outperforming manual training courses designed by humans. https://eureka-research.github.io/eurekaverse/

4. Multi-Task Interactive Robot Fleet Learning with Visual World Models https://ut-austin-rpl.github.io/sirius-fleet/

5. “AI systems are starting to outperform experts on test questions in areas like physics, chemistry, and coding. Consider GPQA, or Google-Proof Question Answering, for example. This benchmark is basically brand new, and is getting crushed now – AI systems can do better than most experts at solving isolated, graduate-level tasks that are not Google-able, even when the experts are given half an hour to solve the problem.” https://milesbrundage.substack.com/p/should-ai-progress-speed-up-slow

6. Sam Altman says in 5 years we will have "an unbelievably rapid rate of improvement in technology", a "totally crazy" pace of progress and discovery. AI will create "many trillions of dollars" of market cap and next year will be a big push for OpenAI into next-generation AI systems. He says the trajectory of AI model capability improvement will continue "for a long time". https://www.youtube.com/watch?v=peg-aX1oii4

7. OpenAI's Head of Strategic Marketing Dane Vahey says the pace of change and OpenAI's product release schedule are accelerating https://openai.com/business/put-ai-to-work-for-marketing-teams/

8. Microsoft AI CEO Mustafa Suleyman says recursively self-improving AI that can operate autonomously is 3-5 years away and might well be "much, much sooner" https://youtu.be/MgrBuYqvxMg?si=S7Tgxz7-10NBz9b7&t=1133

9. Ex-Google CEO Eric Schmidt says in 5 years, AI systems will be able to write and improve on their own code leading to recursive self-improvement and humans are not ready https://youtu.be/cfbD9bsPlFQ?si=vMQpqPHzmO9N0s4o&t=855

10. Google's homegrown cyberdefense agent finds a real-world vulnerability https://googleprojectzero.blogspot.com/2024/10/from-naptime-to-big-sleep.html

11. Grafana and NVIDIA are working on a large language model for observability, apparently given the awkward name LLo11yPop. The model aims to answer natural language questions about system status and performance based on telemetry data. https://developer.nvidia.com/blog/optimizing-data-center-performance-with-ai-agents-and-the-ooda-loop-strategy/

12. NVIDIA has introduced an AI Blueprint that enables developers to create visual AI agents capable of analyzing and summarizing large volumes of video and image content https://blogs.nvidia.com/blog/video-search-summarization-ai-agents/

13. PatternBoost: Constructions in Mathematics with a Little Help from AI — PatternBoost is a new protocol for using transformers to generate interesting mathematical examples -- say, graphs with many edges and no 4-cycles. https://arxiv.org/abs/2411.00566

14. Visa Has Deployed Hundreds of AI Use Cases. It’s Not Stopping. https://www.wsj.com/articles/visa-has-deployed-hundreds-of-ai-use-cases-its-not-stopping-4febe1b4 [no paywall: https://archive.is/DQkMa]

15. Big study of 187k developers using GitHub Copilot: AI transforms HOW we work. Coders can focus. They do more coding and less management. They need to coordinate less, working with fewer people. And they experiment more with new languages, which would increase earnings $1,683/year. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5007084

16. Rabbit AI is focusing on creating autonomous AI agents capable of performing tasks with minimal human intervention https://www.rabbit.tech/research/a-peek-into-rabbit-s-progress-with-LAM-playground

17. Multi-expert Prompting Improves Reliability, Safety, and Usefulness of Large Language Models https://arxiv.org/abs/2411.00492

18. “This work analyzes how weight decay impacts transformer attention layers, finding that it reduces matrix rank, which can harm language model performance.” https://arxiv.org/abs/2410.23819

19. OS-ATLAS: A Foundation Action Model for Generalist GUI Agents https://arxiv.org/abs/2410.23218

20. Waymo explores using Google’s Gemini to train its robotaxis https://www.theverge.com/2024/10/30/24283516/waymo-google-gemini-llm-ai-robotaxi

21. Meta is now allowing the Llama series of models to be used for national security applications by both US agencies and contractors, including Anduril and Palantir. https://about.fb.com/news/2024/11/open-source-ai-america-global-security/

22. Chinese researchers develop AI model for military use on back of Meta's Llama https://www.reuters.com/technology/artificial-intelligence/chinese-researchers-develop-ai-model-military-use-back-metas-llama-2024-11-01/ [no paywall: https://archive.is/K1CYg]

23. Google launched Learn About. It’s an experimental learning tool that turns search queries into interactive educational experiences https://learning.google.com/experiments/learn-about/signup

24. OpenAI in Regulator Talks to Become For-Profit Company https://www.bloomberg.com/news/articles/2024-11-04/openai-in-talks-with-california-to-become-for-profit-company [no paywall: https://archive.is/3R3EI]

Compute:

1. Parallel and Multiplexed: The New Wave of All-Optical Logic Operations https://spie.org/news/paralleled-and-multiplexed-all-optical-logic-operation

2. Large-scale programmable logic array achieves complex computations https://spie.org/news/large-scale-programmable-logic-array-achieves-complex-computations

3. Tokyo University of Science sets pace in neural networks on edge IoT https://www.computerweekly.com/news/366614778/Tokyo-University-of-Science-sets-pace-in-neural-networks-on-edge-IoT

Space:

1. China plans to crash a spacecraft into a distant asteroid https://www.economist.com/science-and-technology/2024/11/05/china-plans-to-crash-a-spacecraft-into-a-distant-asteroid [no paywall: https://archive.is/AHmh2]

2. Researchers spot black hole feeding at 40x its theoretical limit https://arstechnica.com/science/2024/11/researchers-spot-black-hole-feeding-at-40x-its-theoretical-limit/

3. “Much more work is needed before this is really a robust result (paper forthcoming, hopefully by the end of the year), but the initial findings are clear: Spacetime discreteness may be observationally detectable in things like quasar luminosities.” https://x.com/getjonwithit/status/1853233148462506478

Technology:

1. Biochemists Take Key Steps Toward Synthetic Lifeforms https://research.rug.nl/nl/clippings/creating-a-simplified-form-of-life-scientists-build-modules-for-s

2. New liquid biopsy method offers avenue to quick, affordable cancer diagnosis https://www.rochester.edu/newscenter/cad-lb-extracellular-vesicles-liquid-biopsy-cancer-diagnosis-624612/

3. “Huawei is expected to install Harmonyos next, its new homemade operating system, on the devices. This would be China’s first clean break with the Western-backed systems on which it and the rest of the world rely.” https://www.economist.com/business/2024/11/05/huaweis-new-made-in-china-software-takes-on-apple-and-android [no paywall: https://archive.is/A2d7a]

Miscellaneous:

1. Survival without dignity (recommended AI-related science fiction featuring the anthropic principle) https://www.lesswrong.com/posts/BarHSeciXJqzRuLzw/survival-without-dignity

2. “We Fell For The Oldest Lie On The Internet” https://www.youtube.com/watch?v=bgo7rm5Maqg

3. “Germanic societies are the most impersonally honest, with other (non-Soviet) Europeans and sometimes Japan behind them. The rest of the world, including high-IQ China, is deeply dishonest as a rule.” https://x.com/arctotherium42/status/1852778703987548215

https://x.com/MoreBirths/status/1852574350005748174

Ukraine:

1. This is what Donald Trump said one year ago: "Before I even arrive in the Oval Office I will have the war between Russia and Ukraine settled. I will get the problem solved in rapid order, it will take me no longer than one day". https://x.com/NOELreports/status/1854110973885571209

2. "Aid to Ukraine" is really an investment in American manufacturing. https://x.com/ColbyBadhwar/status/1853440143522205931

3. “The lessons from the ongoing wars in Ukraine and the Middle East are straightforward and enduring: mass matters, and even the most exquisite system deployed in small numbers with too few munitions will ultimately fail to overcome attrition.” https://www.realcleardefense.com/articles/2024/11/04/the_pentagon_should_rethink_what_weapons_and_munitions_it_should_buy_1069621.html

4. Drone strike on Russian port in Kaspiysk. ~1000km from the frontline. https://x.com/bayraktar_1love/status/1854080114943762761 (The moment the Ukrainian drone hits its target in Kaspiysk, Russian marines make their way out. https://x.com/NOELreports/status/1854100810822525372)

5. “After each drop, dust rises from every russian invader — that's the impact of metal pellets.” https://x.com/GloOouD/status/1852681539294814288

6. Long video from Madyar showing FPV, Vampire, and UAV-dropped munition strikes from October on Russian tanks, BTR-82AT, BMP-3, motorcycles, MRAPs, Desertcross ATVs, towed howitzers, buildings, antenna, and defensive positions. https://x.com/RALee85/status/1852837374310445526

7. “The Siversk direction is heating up. Recently, the 10th brigade repelled a powerful Russian assault. This time, they teamed up with the 3rd Border Detachment. The Russian failed attack resulted in the loss of 4 BMPs and ~20 assault forces.” https://x.com/NOELreports/status/1853786515693858989

8. “Russian channels are highlighting the disastrous events of November 2, when Ukraine's 10th Mountain Assault Brigade inflicted severe losses on a coordinated push by four battalions in the Bilohorivka area, along the Siversk axis. According to Russian sources, it was the reckless order of the 123rd Brigade commander that led to this incident, resulting in significant, irrecoverable losses in both personnel and equipment.” https://x.com/wartranslated/status/1854105773816615322

9. “A speech by Alexander Borodai, a member of the Russian parliament and a participant in the war in Donbas since 2014, has leaked online, addressing the so-called Russian volunteers. He openly referred to them as "second-rate soldiers" and "surplus people," used solely to exhaust Ukrainian defenses. According to him, the hundreds of thousands joining the Russian army are ballast, incapable of productive activity.” https://x.com/wartranslated/status/1853378662478873059

10. “Ruins of the 350-year-old city of Vovchansk, completely destroyed by the so-called Russian army during their unsuccessful offensive in the Kharkiv region. Just six months ago, this thriving town was home to thousands of people.” https://x.com/wartranslated/status/1853046385949524182

11. Over half of foreign respondents (54%) want Ukraine to win the war, while 20% prefer Russia, according to a survey by The Economist. The poll included 30,000 people from 29 countries and Hong Kong. Solidarity with Ukraine was strongest in Europe and among U.S. allies like South Korea and Japan. https://www.economist.com/international/2024/11/03/what-the-world-thinks-of-trump-ukraine-and-chinese-supremacy [no paywall: https://archive.is/E9ME3]