In a podcast interview, Sam Altman disclosed that India has overtaken the US as OpenAI's largest market by user count, attributing growth to strong demand and scarce supply in fast-growing economies. On the wave of senior departures, he framed the split as a belief that 'only I can be trusted to do this,' while conceding that OpenAI holding the technology alone would not be a good outcome.
Scaling Laws
缩放定律
Empirical rules describing how model performance improves with increased parameters, data, and compute.
Curated Intelligence
Turing Award winner Richard Sutton, author of the influential 2019 essay "The Bitter Lesson," argues in a new post that raw compute alone is insufficient for intelligence and that AI science needs clever new methods and is far from finished. He has previously criticized the limits of LLM scaling, called synthetic data a "big mistake," and called for agents with world models that continuously learn from their own experience.
Even as GPT-6 Astra's launch prompted Jensen Huang to declare that AGI has arrived, co-creators of today's models are leaving OpenAI, Google DeepMind and Meta to found new labs that raise billions with no product. The piece sorts 45 such neolabs by research bet, maps the limitations of current AI they aim to overcome, and argues they are wagering that the existing recipe alone will not be enough.
OpenAI's Ted Sanders notes that while LLM performance is multidimensional, model size remains the primary driver of reliability, cost, and recall. He also clarifies that version numbers reflect release order, training data similarity, and the magnitude of improvement.
Former Google CEO Eric Schmidt, historian Niall Ferguson, and others argue in Foreign Affairs that OpenAI and Anthropic resemble modern colonial trading companies. They contend that AI's reliance on compute, energy, and capital creates a 'plantation economics' dependency on the US and China, proposing an AI détente to cap dangerous capabilities.
The article explores applying the 'Scaling Hypothesis' that drove LLM progress to robotics. It highlights that while language models achieved leaps via massive data, robotics faces significant hurdles in gathering comparable real-world interaction data, with the field currently in an early stage of validating this paradigm.
Pushmeet Kohli from Google DeepMind and Sal Candido from Biohub discuss the future of AI in biology. They argue that AlphaFold's structural prediction is merely a starting point, as scaling compute alone cannot solve all protein folding mysteries. The panel highlights the need to move beyond individual proteins toward cellular system modeling and discover new scaling laws in biological data.
AWS and Hugging Face released a guide on foundation model training and inference architecture, emphasizing the shift from raw compute scaling to efficient distributed infrastructure optimization. The article reviews Kaplan's scaling laws and details specific technical components for optimizing large-scale accelerator utilization and distributed training on AWS. It primarily targets enterprise users, offering best practices for cloud-native AI infrastructure.
Hugging Face published a technical blog detailing the implementation and evolution of Mixture of Experts (MoE) architectures within the Transformers library. The article reviews the paradigm shift from dense to sparse models, exploring recent advances in training efficiency and inference optimization, as well as future abstraction designs.
The ICLR 2026 paper 'ATLAS' presents the largest public multilingual pre-training study to date, covering 774 runs across 400+ languages. It introduces adaptive transfer scaling laws that quantify synergies between 1,400 language pairs, providing data-driven guidance on efficiency and cost for building multilingual models.