Wednesday / AI 101 series: Your guide to Physical AI

Friday / Interview: Spenser Huang from Nvidia

Platforms win by scale. More users make the system better. China has a huge user base that improves US AI platforms through volume and usage frequency.

App ecosystems matter. Like phones need apps, AI hardware needs many models. Letting Chinese developers build and run models on US chips grows the CUDA/model ecosystem and strengthens the architecture globally.

Set the global standard. If American tech is available in China, it’s more likely to become the worldwide default stack – benefiting both US industry and national security.

Developers are the leverage. A large share of the world’s AI researchers are in China; winning their loyalty keeps the American stack ahead.

The U.S. must expand energy supply to win the AI race. Pro-energy, pro-growth policy is essential for any industrial buildout.

Use every form of energy. AI is important enough that all sources should be deployed. We need all the energy to win.

But the most important is the win on efficiency. Grace Blackwell delivers ~10× the performance of Hopper at ~50% more power, so system-level energy efficiency improves sharply.

Strategy: pair more energy with better efficiency rather than relying on brute-force energy alone.

US strikes $80 billion deal for new nuclear power plants by Reuters

The Continual Learning Problem by Jessy Lin

Jensen Huang’s Keynote

NVIDIA

• NVIDIA ARC and Partnership with Nokia for 6G
  A new product line, NVIDIA Arc Aerial RAN Computer, was unveiled – a fully software-defined, programmable computer that merges wireless communication and AI processing on the same platform. In a landmark move, Nokia will integrate ARC into its future base stations.
  Why it matters: “AI on RAN” – AI on the Radio Access Network – is a new kind of edge cloud, where industrial robots, drones, and cars can think locally and act instantly. Very impressive.

• The Rubin Platform and the Future of AI Supercomputing
  Cableless, liquid-cooled, and absurdly powerful – Rubin, NVIDIA’s third-generation, rack-scale supercomputer, has 100 petaflops in a single rack. That’s a 100x the power of the DGX-1 delivered to OpenAI just nine years ago.
  Why it matters: Rubin shows how NVIDIA plans to outgrow Moore’s Law through “extreme co-design.” Huang loves this concept of “co-design.” The company now does scale up (the chips), scale out (the racks), and scale across (AI factories). This is achieved by building racks into a single giant GPU, connecting those racks with AI-optimized networking, and then linking entire data centers together. He calls it “the virtual cycle of AI”. With half a trillion dollars in cumulative orders already visible through 2026, Huang’s message was unmistakable – the AI industrial age has arrived.

• Omniverse DSX – Digital Twins for AI Factories
  Huang also introduced NVIDIA Omniverse DSX, a full-stack blueprint for designing and operating gigascale AI factories. Everything from cooling systems to compute clusters can be co-designed virtually before a single wall is built.
  Why it matters: Building an AI factory is no small feat – it’s energy, physics, and capital at planetary scale. With DSX, the entire process can be simulated and optimized beforehand. It’s the digital twin of industrial civilization, reducing years of planning to months.

• NVIDIA Drive Hyperion and Partnership with Uber
  Finally, NVIDIA Drive Hyperion – the standardized architecture for “robotaxi ready” vehicles. Already adopted by Mercedes-Benz and Lucid, it now joins forces with Uber. Hyperion cars will plug directly into Uber’s global network.
  Why it matters: This is the missing bridge between self-driving prototypes and commercial deployment. By unifying sensors, compute, and infrastructure, NVIDIA is turning autonomous driving from a moonshot into a market. Incredible range of announcements.

OpenAI: Unchained Ambition
OpenAI is now officially off the nonprofit leash. A new restructuring deal with Microsoft removes prior fundraising constraints, positioning the ChatGPT maker for a likely IPO. Microsoft retains a 27% stake, but forfeits exclusivity on compute. Altman’s vision includes $1.4 trillion in data centers and gigawatt-scale ambitions. The nonprofit still holds the reins – at least on paper. Microsoft’s return? Nearly 10x on its $13.8B bet.

Google: Entanglement Meets Advantage

Google and Anthropic (also – follow us on Twitter :)

Every attention matters and Every step evolves
Researchers from Ling Team, Ant Group developed Ring-mini-linear-2.0 (16B parameters, 957M activations) and Ring-flash-linear-2.0 (104B parameters, 6.1B activations), using a hybrid architecture combining linear and softmax attention. These LLMs cut inference costs to 1/10 compared to 32B dense models and over 50% vs. earlier Ring models. Training efficiency improved 50% via the custom FP8 library linghe. Optimal attention ratios were identified, enabling SOTA performance in complex reasoning benchmarks →read the paper They also released Ring-1T, a 1-trillion parameter Mixture-of-Experts model activating 50B parameters per token. To tackle training-inference misalignment and RL inefficiencies, they introduced IcePop (token-level discrepancy masking), C3PO++ (dynamic rollout partitioning), and ASystem (a high-performance RL framework). Ring-1T achieves 93.4 on AIME-2025, 86.72 on HMMT-2025, 2088 on CodeForces, 55.94 on ARC-AGI-1, and silver medal level on IMO-2025, setting new open-source reasoning benchmarks →read the paper

DeepAnalyze: Agentic large language models for autonomous data science
Researchers from Tsinghua University introduced DeepAnalyze-8B, an 8B-parameter agentic LLM that autonomously generates analyst-grade research reports from raw data. It uses a curriculum-based training approach simulating human learning, and a data-grounded trajectory synthesis framework for high-quality training data. DeepAnalyze handles tasks from data Q&A to open-ended analysis. Despite its size, it outperforms workflow-based agents built on proprietary LLMs. All model components, including code and data, are open-sourced →read the paper

Aion-1: Omnimodal foundation model for astronomical sciences
Researchers from Flatiron Institute, Berkeley and other institutions introduced AION-1, a multimodal foundation model for astronomy integrating imaging, spectroscopic, and scalar data using modality-specific tokenization and transformer-based masked modeling. Pretrained on over 200 million observations from surveys like SDSS, DESI, and Gaia, AION-1 (300M–3.1B parameters) excels in tasks such as galaxy property estimation, morphology classification, and spectral super-resolution. It uses a single frozen encoder and is fully open-sourced, offering a scalable blueprint for scientific multimodal modeling →read the paper

BAPO: Stabilizing Off-Policy Reinforcement Learning for LLMs via Balanced Policy Optimization with Adaptive Clipping – stabilize off-policy RL by adaptively balancing positive and negative gradients, preserving entropy and exploration for data-efficient, large-scale reasoning improvements →read the paper

LoongRL: Reinforcement Learning for Advanced Reasoning over Long Contexts – train models to reason across long contexts by generating complex multi-hop tasks that teach retrieval, planning, and rechecking at scale →read the paper

🌟 QueST: Incentivizing LLMs to Generate Difficult Problems (by University of Zurich, Microsoft)) – generate high-difficulty coding problems through difficulty-aware sampling and rejection tuning to advance reasoning and distillation data quality →read the paper

🌟 Unified Reinforcement and Imitation Learning for Vision-Language Models (by NVIDIA, Kaist) – combine RL with adversarial imitation to help small VLMs mimic and improve upon teacher models across modalities →read the paper

Search Self-play: Pushing the Frontier of Agent Capability without Supervision – achieve self-supervised RL for search-based agents by co-evolving task proposers and solvers in retrieval-augmented self-play →read the paper

🌟 On-policy Distillation (Thinking Machines) – train smaller models by sampling their own outputs and receiving token-level feedback from a teacher, achieving RL-level performance with 10–30× less compute →read the paper

🌟 Deep Self-Evolving Reasoning – model reasoning as a stochastic process where self-improvement probability compounds across iterations, enabling small models to surpass their teachers through long-horizon refinement →read the paper

🌟 Efficient Long-context Language Model Training by Core Attention Disaggregation – decouple attention into dedicated compute servers to balance workloads and eliminate stragglers, achieving up to 1.35× throughput on 512 GPUs →read the paper

🌟 Glyph: Scaling Context Windows via Visual-Text Compression (by Zhipu AI) – render text as images for visual-language processing, compressing long contexts 3–4× while maintaining reasoning accuracy and enabling million-token tasks →read the paper

🌟 LightMem: Lightweight and Efficient Memory-Augmented Generation – organize LLM memory into sensory, short-term, and long-term layers to preserve past context efficiently with over 10× runtime savings →read the paper

AdaSPEC: Selective Knowledge Distillation for Efficient Speculative Decoders – accelerate speculative decoding by filtering difficult tokens during distillation, improving acceptance rates and generation efficiency →read the paper

ARC-Encoder: Learning Compressed Text Representations for Large Language Models – train portable encoders that replace token embeddings with continuous compressed representations to reduce context length and inference cost →read the paper

Sparser Block-Sparse Attention via Token Permutation – reorder tokens to maximize block sparsity and achieve up to 2.75× faster prefilling while maintaining accuracy close to full attention →read the paper

DeepAgent: A General Reasoning Agent with Scalable Toolsets – unify reasoning, tool discovery, and action execution with RL-based tool credit assignment and dynamic memory folding for scalable, autonomous agents →read the paper

Thought Communication in Multiagent Collaboration – propose “thought communication,” enabling agents to exchange latent states directly instead of language, revealing theoretical guarantees for shared and private thoughts →read the paper

🌟 Extracting Alignment Data in Open Models (by Google DeepMind) – show that alignment datasets can be semantically extracted from post-trained models using embedding similarity, revealing risks and implications for distillation practices →read the paper

Soft Instruction De-escalation Defense – mitigate prompt injections with iterative sanitization and rewriting loops for safer tool-augmented LLM agents →read the paper

Language Models are Injective and Hence Invertible – prove that transformers are injective mappings, allowing exact reconstruction of inputs from activations, enabling theoretical and practical interpretability →read the paper

RECALL: REpresentation-aligned Catastrophic-forgetting ALLeviation – merge models via hierarchical, representation-based fusion to prevent forgetting and align domain knowledge without old data →read the paper

From Masks to Worlds: A Hitchhiker’s Guide to World Models – trace the evolution from masked models to interactive, memory-augmented architectures, charting the path toward persistent generative world models →read the paper

Redefining Retrieval Evaluation in the Era of LLMs – propose utility- and distraction-aware metrics (UDCG) that correlate better with generation quality than classical IR measures →read the paper

A Definition of AGI – define AGI quantitatively through ten cognitive domains based on psychometric theory, establishing measurable cognitive profiles and AGI scoring for modern models →read the paper Unfortunately this paper evoke some critique: