xynova

𝗼𝟯'𝘀 𝗡𝗶𝘁𝗿𝗼𝘂𝘀 𝗕𝗼𝗼𝘀𝘁 𝘁𝗼 𝗕𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝘀 🚀 This year seems to be closing with lots of releases in the world of AI, and specially a big announcement from OpenAI, which are coming Fast And Furious with a new o3 model. 📈 The model demonstrates qualitative performance improvements across various benchmarks, offering interesting insights into potential AI developments in the coming year. https://lnkd.in/gA9idVW9 🏆 For me , two of the most interesting scores are: - On ARC-AGI: o3 more than triples o1's score on low compute and surpasses a score of 87% - On EpochAI's Frontier Math: o3 set a new record, solving 25.2% of problems, where no other model exceeds 2% 🤖 The ARC organizers are impressed with o3's capabilities, stating it's "capable of adapting to tasks it has never encountered before, arguably approaching human-level performance in the ARC-AGI domain." While they emphasize that o3 is not yet AGI and will face significant challenges with the upcoming ARC-AGI-2 benchmark, they view its progress as more than incremental. They assert that o3 represents "new territory" in AI development, warranting serious scientific attention. 💰 Cost also remains a consideration, with each problem in low-compute mode priced at $17-20. However, as we've seen historically, compute costs are likely to decrease, potentially bringing this technology within reach of consumer-grade hardware in the coming years. 💡 A key innovation of o3 is its ability to "fix the fundamental limitation of the LLM paradigm – the inability to recombine knowledge at test time," achieved through "LLM-guided natural language program search." For more details, you can refer to their blog post: https://lnkd.in/g66_TSXs ✨ It feels like Sam Altman's "There is no wall" assertion stands true as we finish the year 2024. ℹ️ About ARC-AGI: ARC-AGI is a benchmark that tests AI systems' ability to solve novel reasoning tasks through visual pattern recognition, designed to evaluate genuine reasoning capabilities rather than pattern memorization or specialized knowledge. ℹ️ About EpochAI's Frontier Math: The benchmark is designed to evaluate an AI's ability to solve novel mathematical problems that require deep reasoning, proof construction, and mathematical intuition rather than just calculation or formula application. https://lnkd.in/g6_U2Tna

Can we really brainwash AI? To understand what AI thinks, you need to ask the right questions and hope it provides truthful answers. However, this approach is not ideal and is closely related to how these models are trained. We essentially feed these models a vast amount of data and constrain the space available for them to internalize this information. Additionally, while the data is being processed, we don't really instruct the models to organize it in a specific way. It's similar to throwing high-speed books at a librarian and asking them to sort the books in any manner they see fit, with the only goal being not to drop any. As a result, these models become uninterpretable, making it impossible to reconstruct a coherent picture of how the information is structured. The model itself doesn't fully understand its own processes; it only knows that there are references to certain information scattered throughout its internal workings. In recent years, significant efforts have been made to make these models more interpretable. One popular approach involves using sparse autoencoders. Explaining this concept can be confusing, so I'll let the video do the talking. Will this be the solution? Probably not. I believe it will need to be scaled up to the same extent as the models themselves to become truly effective. However, it might trigger the development of new architectures or innovative ways of training these models. Enjoy the video. https://lnkd.in/g_tRaDDx and a playground: https://lnkd.in/giPsk9H2

🌪️ Denoising towards the eye of the storm: Diffusion architecture in weather forecasting? Google DeepMind has unveiled GenCast, quite a novel AI system that employs diffusion architecture to generate weather forecasts up to 15 days in advance. This innovative approach adapts techniques traditionally used in image, video, and sound generation to the domain of weather prediction. Diffusion models work through an iterative refinement process, starting with a noise-initialized state and gradually refining it into a coherent prediction. In GenCast's case, this process is applied to weather forecasting, where the model essentially "denoises" atmospheric states to produce realistic weather scenarios 🤯. What is really interesting about this approach is that it bypasses the traditional resource intensive physical simulation technique and takes a "creative probabilistic shortcut" that shows some hard to believe performance: ✨ It can produce a complete 15-day global forecast in approximately 8 minutes using a single Google Cloud TPU v5, a big contrast to the resource-intensive supercomputer requirements of traditional methods. 🎯 It outperforms ECMWF's ENS (the current top operational ensemble forecast) on 97.4% of evaluated targets. 📊 Its forecast probabilities reflect the actual occurrences with greater accuracy. ⚡ It shows particular skill in forecasting extreme weather events and tropical cyclones. It is important to note that GenCast still relies on traditional numerical weather prediction (NWP) systems for initial condition setup. However, this hybrid approach could pave the way for a simpler, more affordable architectures based on probabilistic modelling. Such architectures could potentially address a wide range of "infinite problem" scenarios, where the solution space is vast and complex. How interesting, don’t you think? DeepMind blog: https://lnkd.in/ea6J8CTv

👀 Cerebras running Big Models at Instant Speeds is something to watch out for. With this company just having filed for an IPO, and inference-time compute being a proven path to push AI models beyond parameter size and data availability, their situation is intriguing. Their ahead-of-its-time "bigger is better" approach to hardware design could play out very interestingly in the coming couple of years (AI Inference Chip Market Size And Forecast: https://lnkd.in/gC6gh9Sc) Youtube: WSE architecture: https://lnkd.in/gaqxht5j A few highlights about their numbers: 🧮 Load massive models up to 24 trillion parameters on one chip 💻 900,000 cores for distributed model weight processing 💾 2.4 Petabytes of high-performance memory (MemoryX) ⚡ 7,000x more memory bandwidth than leading GPUs 🚀 Higer speed performance during both training and inference time 🌱 Lower power consumption and reduced infrastructure complexity 💰 Competitive pricing: Cerebras Inference at 10¢ per million tokens - https://lnkd.in/gYpEBvUt World Record Meta's Llama 3.1-405B Model clocking 969 Tokens per Second with 240ms latency - https://lnkd.in/e4WFqh49 This new world record means that scientists can now complete two years’ worth of GPU-based simulation work every single day

🤔 Could AI be used one day to eliminate social issues like Bullying and Gender Bias? The National University of Singapore's paper "Multi-Agents are Social Groups" reveals how AI agents can shape human opinions! 🧠💡https://lnkd.in/dcEBSnEc 1️⃣ Artificial Social Norms: Multiple AI agents create a sense of agreement, forming social norms! 👥👍 2️⃣ Psychology at play: • Agents use "I" statements for independent thoughts 🗣️ • Multiple agents echoing opinions = perceived consensus 🔊🔊🔊 3️⃣ Magic number is 3️⃣: • Perfect for consensus without skepticism 🎯 • More powerful than solo agents 💪 🔍 Here's a thought: Could democracy drive the behavior of these multi-agents to promote positive change? Change that is democratically voted by the people themselves? 🤯🤯🤯🤯 Youtube > Multi AI Agent System - Pure Social Manipulation: https://lnkd.in/gy8Xi5vg

Automated ML researchers are coming soon to lower the entry barriers to experimentation: Exciting times!. NEO A fully autonomous Machine Learning Engineer "When put to the test across 50 Kaggle competitions, NEO didn't just participate — it excelled, securing medals in 26% of the competitions beating the OpenAI's benchmarks. To put this achievement in perspective, earning a gold medal in Kaggle requires performing in the top 10% of all participating teams — a feat that typically demands exceptional expertise, innovative approaches, and meticulous optimization." https://heyneo.so/blog

Did you know most of Nvidia's high-end chips, particularly those used for AI, are manufactured in Taiwan by TSMC? 🤔 No wonder TSMC is rushing to set up shop in the USA! When 92% of America's advanced chips come from an island China's eyeing 👀, you'd be in a hurry too. 🏃♂️➡️ It makes you think how fragile is our tech supply chain really? 🤔💭 Youtube: https://lnkd.in/d3v49YFT

How interesting is this? MindsAI researchers developed a model that can learn and adapt in real-time, adjusting to new tasks even with limited prior experience, as part of their ARC challenge solution. Their model achieved a score of 55% on the ARC benchmark, where 85% is considered equivalent to human-level intelligence. 🔄 Pre-training: - They first fine-tune a language model on a large dataset of custom-generated ARC-like tasks - This helps the model understand the general structure and patterns of ARC problems 🧠 🎯 Active Inference: - During testing, they perform additional fine-tuning on each specific task - This allows the model to adapt its knowledge to the unique aspects of each individual problem ✨ About the ARC score: An ARC (Abstraction and Reasoning Corpus) score of 85% is considered a key benchmark for AGI (Artificial General Intelligence), representing human-level performance on abstract reasoning tasks and demonstrating significant progress toward general intelligence capabilities. This threshold indicates several important capabilities: the ability to identify and apply abstract patterns, flexible problem-solving across novel scenarios, potential for human-comparable reasoning abilities, and represents a major milestone in AI development 🤖 🏆 Youtube: https://lnkd.in/gUu8i8RK

Imagine if during the Cold War, the Soviet Union had taken an American invention, modified it for military use, and then publicly announced it to the world. 🌶️🔥 Reuters: https://lnkd.in/ehcsg2Y8

"AI is moving too fast, and is too complex, for us to rely exclusively on a small cohort of large firms; we need to empower and learn from a full range of talented individuals and institutions who care about making AI safe, secure, and trustworthy." https://lnkd.in/eNG56Y_G