Links for 2024-10-31
AI:
🌐 Introducing ChatGPT search 🌐: “ChatGPT can now search the web in a much better way than before so you get fast, timely answers with links to relevant web sources. The search model is a fine-tuned version of GPT-4o, post-trained using novel synthetic data generation techniques, including distilling outputs from OpenAI o1-preview.” https://openai.com/index/introducing-chatgpt-search/
Centaur: a foundation model of human cognition — “Centaur not only captures the behavior of held-out participants better than existing cognitive models, but also generalizes to new cover stories, structural task modifications, and entirely new domains. Furthermore, we find that the model's internal representations become more aligned with human neural activity after finetuning.” https://arxiv.org/abs/2410.20268
Can Graph Learning Improve Planning in LLM-based Agents? Novel algorithm with strong performance (even doubling the performance of GPT) https://arxiv.org/abs/2405.19119
Relaxed Recursive Transformers: Effective Parameter Sharing with Layer-wise LoRA https://arxiv.org/abs/2410.20672
The Geometry of Concepts: Sparse Autoencoder Feature Structure https://arxiv.org/abs/2410.19750
IRIS — A new approach called IRIS combines large language models (LLMs) with static analysis to detect security vulnerabilities in software. Using a dataset called CWE-Bench-Java, IRIS detected 69 out of 120 vulnerabilities in Java projects, outperforming traditional static analysis tools that found only 27. https://arxiv.org/abs/2405.17238
Sundar Pichai said on the earnings call today that more than 25% of all new code at Google is now generated by AI. https://blog.google/inside-google/message-ceo/alphabet-earnings-q3-2024/#full-stack-approach
OpenAI CFO Sarah Friar says AGI is "closer than most think" and the ability of internal research models to perform at PhD level in a range of fields "would blow your mind to see what's coming" https://www.youtube.com/watch?v=eCqFgVqWbEs
Elon Musk says AI is improving at the rate of at least 10x per year and will be able to do anything a human can do in a year or two and be equal to the intelligence of all humans combined 3 years after that https://youtu.be/3JkkWfzc4Jg?si=LM5YDFRvkZ3yPx0D&t=75
Sam Altman tells the OpenAI's London DevDay that their o series of reasoning models are "on a quite steep trajectory of improvement" and "I would encourage people to be aligned with that" https://youtu.be/VTeRZqUHi4E?si=s4lzfq5v573u0qi0&t=50
Bill Gates says AI will be able to model human biology and health beyond our ability to comprehend and it will lead to many health problems being solved in the next 10-20 years https://youtu.be/KeGYI69sWvw?si=mGoRBV6vE730aU6Y&t=2051
Once seen as a speculative bubble, AI companies “are likely to continue driving returns for investors,” an analyst at Goldman Sachs wrote. https://www.goldmansachs.com/insights/articles/ai-stocks-arent-in-a-bubble
Podcast that dives into SWE-bench, SWE-agent, and most recently SWE-bench Multimodal with John Yang from Stanford University and Carlos E. Jimenez from Princeton University https://www.youtube.com/watch?v=8rwHAR4fsFg
“How Do We Build a General Intelligence? 1) How do we build systems that learn and generalize, from a perspective of probability and compression? Can we use these principles to resolve mysterious generalization behaviour in deep learning? 2) Is it possible to build general-purpose AI systems in light of results like the no free lunch theorems? 3) What are the prescriptions for general intelligence? 4) What are the demonstrations of those principles in scientific settings? 5) What are we far away from solving?” https://www.youtube.com/watch?v=HEp4TOrkwV4
Beyond Text: Optimizing RAG with Multimodal Inputs for Industrial Applications https://arxiv.org/abs/2410.21943
Robotics:
Physical Intelligence (π) show the first step towards bringing general-purpose AI into the physical world. The first generalist model π₀ 🧠 🤖 https://www.physicalintelligence.company/blog/pi0
Meta cutting-edge developments in robotics and touch perception: 1️⃣ Meta Sparsh is the first general-purpose encoder for vision-based tactile sensing that works across many tactile sensors and many tasks. Trained on 460K+ tactile images using self-supervised learning. 2️⃣ Meta Digit 360 is a breakthrough artificial fingertip-based tactile sensor, equipped with 18+ sensing features to deliver detailed touch data with human-level precision and touch-sensing capabilities. 3️⃣ Meta Digit Plexus is a standardized platform for robotic sensor connections and interactions. It provides a hardware-software solution to integrate tactile sensors on a single robot hand and enables seamless data collection, control and analysis over a single cable. https://ai.meta.com/blog/fair-robotics-open-source/
HOVER: Versatile Neural Whole-Body Controller for Humanoid Robots https://hover-versatile-humanoid.github.io/
Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning https://www.youtube.com/watch?v=GuD_-zhJgbs [project site: https://hil-serl.github.io/]
World models, also known as world simulators, are being touted by some as the next big thing in AI. https://techcrunch.com/2024/10/28/what-are-ai-world-models-and-why-do-they-matter/
moreover, the dynamics model is a whole body model, which enables the robot to do squat motions. This is incredible, no one else on this earth has acheived this.
Thread: https://x.com/ChongZitaZhang/status/1851659956660900024
Compute:
Accelerating AI Performance using Anderson Extrapolation on GPUs https://arxiv.org/abs/2410.19460
Inside xAI's Colossus supercluster, the largest AI supercomputer in the world, with over 100,000 GPUs, exabytes of storage and liquid cooling https://www.youtube.com/watch?v=Jf8EPSBZU7Y (Also: “This video will be outdated in less than two months from now once we bring online an additional 100k GPUs, transforming it into a 200k GPU cluster.” https://x.com/rudaykumarraju/status/1851083357301698931)
META CEO: "We're training the Llama 4 models on a cluster that is bigger than 100,000 H100s, or bigger than anything I've seen reported for what others are doing" https://finance.yahoo.com/news/q3-2024-meta-platforms-inc-140416652.html
OpenAI builds first chip with Broadcom and TSMC, scales back foundry ambition https://www.reuters.com/technology/artificial-intelligence/openai-builds-first-chip-with-broadcom-tsmc-scales-back-foundry-ambition-2024-10-29/ [no paywall: https://archive.is/BKqev]
Cerebras Systems achieves record-breaking 2100 tokens/sec inference speed with Llama 3.1-70B model. https://cerebras.ai/blog/cerebras-inference-3x-faster
Running an LLM on a small customizable chip https://learnandburn.ai/p/running-an-llm-on-a-small-customizable
Biden admin finalizes order restricting AI, chips investment in China https://www.semafor.com/article/10/28/2024/biden-admin-finalizes-order-restricting-ai-chips-investment-in-china
Fab Whack-A-Mole: Chinese Companies are Evading U.S. Sanctions https://www.semianalysis.com/p/fab-whack-a-mole-chinese-companies
Neuroscience:
The Clinically Blind See Again With an Implant the Size of a Grain of Salt https://singularityhub.com/2024/10/28/the-clinically-blind-see-again-with-an-implant-the-size-of-a-grain-of-sand/
Elon Musk says at high volume, Neuralink should approach the cost of an Apple watch or phone and be implanted by a robot in a 10-minute surgery https://youtu.be/huxf36QKbI0?si=cLf0pCoyBCG65nqy&t=1703
Miscellaneous:
Science Is Finding Ways to Regenerate Your Heart https://www.wsj.com/health/grow-heart-lung-tissue-medical-technology-24b22bb4 [no paywall: https://archive.is/hluZP]
Politics:
"Surprisingly low IQs in developing countries", we don't want to know, thanks https://www.emilkirkegaard.com/p/surprisingly-low-iqs-in-developing
Why Is Most Journalism About IQ So Bad? https://quillette.com/2024/10/30/why-is-most-journalism-about-intelligence-so-bad/ [archived version: https://archive.is/TEMdF]
“This is the one pro-Trump argument that genuinely bothers me, but I have four counterarguments.” https://www.astralcodexten.com/p/acx-endorses-harris-oliver-or-stein
Russian state TV: "Trump can really get it to the point that our geopolitical adversary will fall apart!" https://www.youtube.com/watch?v=vkBKuymal8M
Ukraine:
S. Korea is not considering direct provision of 155-mm artillery shells to Ukraine: presidential office https://en.yna.co.kr/view/AEN20241030007400315
“President Zelensky announced during an interview with South Korean TV channel KBS that Ukraine will soon formally request military aid from South Korea.” https://x.com/NOELreports/status/1852015745451852006
“The Russians recently launched a large offensive in eastern and southern Donetsk, on a 70 kilometers wide front. The attack has breached Ukrainian defences in just a few days in many areas, and there can be some dangerous developments ahead, which I’ll discuss in this thread.” https://x.com/emilkastehelmi/status/1851361095329493405
“On manpower, too, Russia remains solvent. Its army is recruiting around 30,000 men per month, says the nato official. That is not enough to meet internal targets, says another official, but it is adequate to cover even the gargantuan losses of recent months.” https://www.economist.com/europe/2024/10/29/ukraine-is-now-struggling-to-survive-not-to-win [no paywall: https://archive.is/1fD7U]
Russian drones hunt civilians, evidence suggests https://www.bbc.com/news/articles/c207gz7key6o
Roads clogged with destroyed Russian armored vehicles somewhere on the front https://x.com/bayraktar_1love/status/1851338655278731473
Thread about Russian losses in Kursk https://x.com/OSINTua/status/1852001740314263563
russian assault, presumably by 810th, with some, interesting dismount behavior. https://x.com/giK1893/status/1851741909519339638
Full video of yesterday’s strike on Russian ammunition storage in Luhansk. https://x.com/bayraktar_1love/status/1851614679983915323
HIMARS strike on Russian BUK air defence system somewhere at the front. https://x.com/bayraktar_1love/status/1852022132869390690