Google put a walking humanoid inside a single policy, published the reasoning half as an API, and kept the motor half behind a partner gate. The numbers explain why.
AI generation is temporarily unavailable; this entry was preserved with deterministic fallback metadata.
Google put a walking humanoid inside a single policy, published the reasoning half as an API, and kept the motor half behind a partner gate. The numbers explain why.
NVIDIA and 24 other firms urge Washington to avoid premature restrictions on open-weight AI; Moonshot open-sources Kimi K3; Google DeepMind unveils Gemini Robotics 2; Leopold Aschenbrenner's Situational Awareness fund sells portfolio after losses; Big Tech earnings reveal uneven AI monetization.
Jensen Huang's first X post backs an open-weight AI letter signed by 25 companies.
Moonshot releases Kimi K3, a 2.8T-parameter MoE model with 1M-token context.
Ilya Sutskever divides AI history into ages of research, scaling, and research again. Yet frontier models still leak Transformers; improvements come from engineering such as data, context, and RL. The author argues breakthroughs stem from the learning loop around the Transformer, not replacement.
Sutskever's periodization is appealing but current models show no architectural revolution.
Improvements in frontier models come from engineering innovations like better data and RL.
This article explores how large language models transform data from a static resource into a transmission medium for intelligence through synthetic data distillation, where a teacher model generates training data for a student model.
Traditional ML treats data as static resource to be mined.
Large language models both consume and generate data like questions, answers, curricula.
Anthropic released Opus 5 advancing long-horizon reasoning and agentic coding; Travis Kalanick's Atoms raised $1.7B for physical AI; Poolside launched Laguna S 2.1 open model; OpenAI models breached safety limits during testing; Alphabet and AMD showcased massive AI infrastructure investments; OpenRouter acquisition rumors highlight distribution layer value.
Anthropic's Opus 5 improves long-horizon reasoning and agentic coding.
Thinking Machine's new model Inkling has 975 billion parameters but activates only 41 billion per token, using a router to select specialist subsets for sparse computation.
Inkling has 975B total parameters but only 41B active per token (about 4.2%).
It uses a sparse activation architecture with a router to choose expert subsets.
This week's developments in AI shift focus from raw scale to distribution and openness. Highlights include Thinking Machines' open-weight Inkling, Moonshot AI's 2.8T parameter Kimi K3, PrismML's phone-runnable Bonsai 27B, OpenAI's self-play red-teaming system GPT-Red, and Xi Jinping's call for open-source AI as a global public good at the World AI Conference in Shanghai.
Thinking Machines released Inkling, a 975B MoE model with open weights and 1M context
Moonshot AI unveiled Kimi K3, a 2.8T parameter model optimized for long-horizon tasks
Meta launched Muse Spark 1.1, the first Meta model with a price tag, marking a shift from open weights to a closed-source business model. As Meta builds a full vertical stack—from chips to cloud to apps—the question arises whether it can compete with frontier AI labs.
Meta released Muse Spark 1.1 with closed weights and paid API, priced at $1.25 per million input tokens and $4.25 per million output tokens, compatible with OpenAI endpoints.
Zuckerberg posted on X for the first time in three years to announce this strategic shift.
OpenAI's audit of SWE-Bench Pro reveals that approximately 30% of benchmark tasks are defective, questioning the validity of precise scores. The finding leads OpenAI to withdraw its recommendation of the benchmark and underscores the need for more reliable evaluation methods.
OpenAI audit finds ~30% of SWE-Bench Pro tasks are flawed
Precise scores can misrepresent model capabilities
Frontier AI labs are shifting from chatbots to integrated systems where models act as runtimes, with near-monthly releases of powerful models and agents. This week's highlights include OpenAI's GPT-5.6 with programmatic tool calling, GPT-Live's full-duplex audio, ChatGPT Work for artifact creation, Meta's Muse Spark 1.1 with active context management, and Grok 4.5 for coding and knowledge work. Research updates reveal issues with coding benchmarks, selective unlearning, agent self-evolution, speculative decoding, and traffic routing. Notable industry news includes major funding rounds for Lovable, Prime Intellect, SambaNova, Norm Ai, and Ollama.
OpenAI releases GPT-5.6 (Sol, Terra, Luna) with programmatic tool calling and parallel subagents.
GPT-Live introduces full-duplex audio interaction, shifting from turn-based to continuous dialogue.
Google Research unveils TabFM, a foundation model for tabular data that performs in-context learning on tables, enabling predictions on unseen datasets with a single forward pass, no training or feature engineering required.
TabFM is a new foundation model for tabular classification and regression from Google Research.
It uses in-context learning to make predictions on entire tables in one pass without training or tuning.
The real history of knowledge distillation predates the famous 2015 Hinton paper by nearly a decade, rooted in 2006 work on compression as mimicry. This article reviews three foundational papers that shaped the field.
Knowledge distillation's history extends back to 2006, with work on compression as mimicry.
The 2015 Hinton et al. paper introduced softmax temperature and 'dark knowledge' but was not the first.
Space is becoming a new frontier for AI, driven by energy scarcity. Low Earth orbit is seen as an unmetered energy source, attracting trillion-dollar companies, chipmakers, and startups. As of December 2025, the first LLM trained in space, nanoGPT, has been completed in orbit.
Space emerges as a competitive AI frontier due to Earth's energy scarcity
Low Earth orbit offers effectively unmetered energy and no jurisdictional zoning
Meta's new Autodata research turns data creation into an agentic process, where models iteratively generate, test, and refine their own training data, shifting the focus from model architecture to data generation.
Autodata treats data generation as an agentic research loop.
AI creates examples, tests them, learns from failures, and updates recipe.
This week in AI saw OpenAI release GPT-5.6 with tiered models and safety architecture, Anthropic introduce Claude Tag for structured collaboration, General Intuition raise $320M for action-model training, and the LayerLens Stratix Cup evaluate models through soccer. Numerous research papers and tech releases also made headlines.
OpenAI unveiled GPT-5.6 suite (Sol, Terra, Luna) with phased access and safety coordination.
Anthropic launched Claude Tag, enabling semantic markers for better model interaction.
Self-driving labs combine AI with automated hardware to let the system learn from experiments and autonomously decide what to do next, moving beyond mere automation to true autonomy.
Self-driving labs use AI to close the loop between design, make, test, and learn.
They differ from automation by making decisions based on real-time results.
One of the main frontier AI models is adding embodied AI capabilities. Alibaba's Qwen-Robot Suite aims to bridge the gap between perception and action with three specialized models.
Qwen models have been confined to software with no physical interaction.
Alibaba launched Qwen-Robot Suite with three models for navigation, manipulation, and world modeling.
A deep dive into one of the most important techniques in modern AI — distillation — and how it addresses the cost, deployment, and specialization challenges of large-scale models.
Distillation makes AI models more efficient and deployable, addressing scale-induced challenges.
Scale drove AI progress but led to expensive, slow, and difficult-to-specialize models.
LayerLens launches the Stratix Cup, a soccer tournament where top AI models compete as agents in a simulated environment, testing planning, adaptation, and multi-agent coordination.
LayerLens introduces the Stratix Cup, a soccer tournament for AI models.
The competition tests agentic capabilities: pre-game strategy, real-time gameplay, and halftime adaptation.
A week of really unexpected turns in the AI market: SpaceX acquires Cursor for $60B, key researchers leave Google, and Midjourney reveals a full-body medical scanner.
1. SpaceX acquires Cursor for $60B in stock, signaling AI tooling as strategic infrastructure.
2. Noam Shazeer and John Jumper leave Google, highlighting talent consolidation in AI frontier.
Google DeepMind has released DiffusionGemma, a text-diffusion model that challenges traditional transformer architectures by not generating text left-to-right token by token.
DiffusionGemma is a text-diffusion model from Google DeepMind.
It challenges the conventional transformer architecture.
This article concludes the series on alternatives to the Transformer, covering four families: recurrent/linear-recurrent models, state space models, text diffusion models, and liquid/continuous-time models. It also announces a new series on knowledge distillation.
Self-attention has quadratic scaling and memory costs for long sequences.
Four alternative directions: recurrent (constant memory), state space (linear scaling), text diffusion (parallel generation), liquid (continuous-time dynamics).