s1: How 1,000 Training Examples Beat OpenAI's o1-preview by 27%
Stanford's s1 model uses 'budget forcing' to exceed o1-preview on math benchmarks with just 1K training examples. The test-time scaling breakthrough explained.
Blog
Insights on GPU infrastructure, AI, and data centers.
South Korea's HBM4 Moment: How Samsung and SK Hynix Became the Gatekeepers of AI
Samsung and SK Hynix control 90% of global HBM production. With HBM4 mass production launching February 2026 and 900K wafers pledged to Stargate, memory has become a strategic weapon. Server DRAM pric...
Singapore Opens 200MW Data Center Allocation with 50% Green Energy Mandate
Singapore's DC-CFA2 allocates 200MW with mandatory 50% renewable energy. Applications close March 31, 2026. AI workloads prioritized. Land-scarce city-state redefines DC standards.
Singapore's Green Data Center Gamble: How 734 Square Kilometers Became AI's Most Competitive Market
Singapore lifted its moratorium with the strictest sustainability requirements in APAC. DC-CFA2 allocates 200MW with 50% green energy mandate. Jurong Island adds 700MW. Vacancy rates hit 1.4%.
Marvell's $540M XConn Acquisition Signals AI Interconnect Consolidation
Marvell acquires XConn for $540M, marking major consolidation in CXL/UALink switching silicon for AI data centers.
China Activates 1,243-Mile AI Computing Network: 98% Efficiency at Continental Scale
China's Future Network Test Facility spans 1,243 miles, achieving 98% single-datacenter efficiency. Xi touts AI breakthroughs. $70B investment wave coming.
Thinking Without Tokens: How Latent Reasoning Makes a 3.5B Model Perform Like 50B
A new architecture lets AI models reason in latent space instead of generating tokens. The recurrent depth approach scales test-time compute without chain-of-thought.
LTX-2: The First Open-Source Model That Generates Synchronized Video and Audio
Lightricks releases LTX-2 with 14B video + 5B audio parameters. Native 4K at 50fps with lip sync, foley, and ambient sound. Fully open weights.
Recursive Language Models: Teaching AI to Manage Its Own Context
MIT's RLM architecture lets models delegate context to sub-LLMs and Python scripts. 100x context extension with 2-3x token efficiency. Prime Intellect predicts the paradigm of 2026.
MiroThinker: The Third Scaling Dimension for AI Agents
MiroThinker introduces interaction scaling—training agents to handle 600 tool calls per task. 81.9% on GAIA benchmark. A new dimension beyond model size and context.
Australia's AI Infrastructure Moment: $100 Billion Bet on Sovereign Compute
OpenAI's $7B Sydney campus, Amazon's $20B AWS expansion, and Google's Christmas Island defense play converge. Australia emerges as APAC's sovereign AI hub with 3.2 GW capacity by 2035.
Microsoft's $60B Neocloud Bet: Buying Time in the AI Capacity Crisis
Microsoft commits $60B+ to neocloud providers including $23B to Nscale for 200K GB300 GPUs. Azure capacity crunch extends into 2026. Neoclouds reshape AI infrastructure.