Xi launches rival AI governance bloc with 29 countries, positioning China as global leader

Moonshot AI's Kimi K3 (2.8T params) achieves frontier performance with 1M-token context, takes #1 on code benchmarks

Moonshot AI released Kimi K3, a 2.8 trillion-parameter mixture-of-experts model with a million-token context window, scoring 57 on Artificial Analysis Intelligence Index (behind Claude at 60, ahead of Opus at 56) and achieving #1 on Frontend Code Arena for the first time with 64% on DeepSWE. The model's Delta Attention mechanism enables 6x faster throughput at 1M context, signaling that efficiency stack (MoE routing, quantization, data curation) may matter more than raw FLOPs for closing the capability gap. This mirrors DeepSeek's pattern of commoditizing premium AI capabilities through open release and challenges the compute-moat narrative that has dominated Western AI strategy.

Meta negotiates ~$10B compute lease to Anthropic; hires AWS infrastructure veteran to lead data center buildout

Meta is in talks to lease approximately $10 billion in excess data center capacity to Anthropic over two years, signaling that Meta's infrastructure is overbuilt relative to near-term internal demand while Anthropic faces acute compute constraints. Concurrently, Meta hired Dave Brown, AWS's former SVP of Compute and AI, to lead its data center expansion—a structural move that could meaningfully improve execution on compute monetization and address past failures in B2B initiatives. Together, these moves suggest Meta is pivoting toward B2B compute provision as a revenue stream while maintaining frontier model development, mirroring SpaceX's approach to monetizing spare capacity.

Thinking Machines' Inkling: open-weight model designed for enterprise fine-tuning, not frontier capability

Mira Murati's $12 billion startup Thinking Machines launched Inkling, an open-weight model positioned as smaller but tunable for specific use cases rather than competing on raw capability. Bridgewater has already deployed a custom version that beat proprietary models on financial reasoning at 1/10th the cost, signaling a strategic shift away from frontier labs' one-size-fits-all approach toward owned, customized intelligence that keeps sensitive data on-premises.

OpenAI's Jony Ive-designed device confirmed as screen-free smart speaker, launching early 2027 at $200–300

Bloomberg confirmed that OpenAI's long-rumored consumer device, designed by Jony Ive, is a personality-driven household companion with camera, microphone, and appliance control rather than a spec-driven gadget, launching in early 2027 at $200–300. Apple's pre-leak lawsuit over alleged trade secrets suggests serious competitive stakes. The device represents OpenAI's bet on distributing AI through consumer hardware rather than software alone.

MemoHarness: agent infrastructure and memory architectures emerge as durable moat as frontier models commoditize

As base model capabilities commoditize and open, community consensus has shifted from base model access to orchestration, tools, and domain-specific scaffolding. MemoHarness, a new agent harness that learns from its own past executions, achieved 0.806 on Shell-Agent benchmarks versus 0.722 for fixed-harness baselines while lowering per-task cost. The system decomposes the harness into six editable control dimensions (context, tool, generation, orchestration, memory, output) and applies case-specific adaptation at test time without retraining. This reflects a broader valuemaxxing-vs-tokenmaxxing transition: as intelligence becomes cheaper and more open, durable competitive advantage shifts to the orchestration layer—prompt construction, tool interfaces, memory management, and workflow topology—rather than the underlying model.

DTCC launches live tokenized-securities trading with 25+ firms; full production launch set for October

The Depository Trust & Clearing Corporation, which clears and settles most U.S. securities trades, began running live production trades of tokenized stocks and Treasurys, with more than two dozen firms participating including JPMorgan, Goldman Sachs, BlackRock, Vanguard, and the New York Stock Exchange. Assets being tokenized include Microsoft and Circle Internet Group shares, major-index ETFs, and U.S. Treasurys of varying maturities. The company called it its largest tokenization production effort by breadth of use cases, asset classes, and participants. Full launch is scheduled for October, marking the inflection point where tokenization becomes operational infrastructure rather than proof-of-concept.

Xi pitches China as leader of new global AI order

Reuters reported on Xi Jinping's speech at the World Artificial Intelligence Conference in Shanghai, positioning China as leader of a competing global AI governance framework. Main takeaways:

• Xi launched the World AI Cooperation Organisation (WAICO) with 29 member countries, framed as rival to U.S.-led "Pax Silica" initiative, signaling institutional competition over AI standards and supply chains
• Chinese open-weight models are rapidly closing capability gap: Moonshot AI unveiled Kimi K3 as world's largest open-weight model by parameter count, same week U.S. pulled Anthropic's frontier models
• Xi made first extended remarks on AI safety, calling for human control and early-warning mechanisms, while promoting open-source AI as "global public good" for developing nations
• Only Kazakhstan listed as joining both U.S. (35 countries) and China (29 countries) initiatives, suggesting limited overlap and competing geopolitical coalitions forming around AI governance

Oozy Intelligence in Slow Time

Venkatesh Rao in Contraptions published an essay titled "Oozy Intelligence in Slow Time". Main takeaways:

• Intelligence too cheap to meter enables deployment on trivial tasks (parts documentation) that wouldn't justify human labor, unlocking utility from previously-abandoned inventory
• AI's 'speed of patience'—ability to process high-detail information at rates humans cannot tolerate—reframes what appears as undifferentiated ooze to humans as structured, processable information to AI systems
• The economic impact is not raw speed but volume: AI can perform tedious, grinding work (bureaucratic detail processing) at scales and durations humans reserve only for high-stakes tasks like taxes, fundamentally changing cost-benefit thresholds for information work
• Connects to broader thesis that AI should be understood through training time (thousands of human-years of learning) and inference-time availability rather than raw computational speed

Kimi K3's Delta Attention Mechanism: 6x Throughput Gains at 1M Context via Fast Weights

sdrzn (technical analysis of Kimi K3 architecture) in X (via sdrzn) published an other titled "Kimi K3's Delta Attention Mechanism: 6x Throughput Gains at 1M Context via Fast Weights". Main takeaways:

• Delta Attention replaces full attention with fixed-size fast weights that retrain per-token on GPU, reducing compute scaling from O(context) to near-constant, achieving 6x faster/cheaper throughput at 1M context while maintaining eval parity with full attention
• Fast weights operate as a second set of trainable parameters running at different speed than frozen 2.8T base weights, effectively creating per-request weight updates during inference
• Architecture enables flat pricing past 200k context (vs. tiered pricing competitors use), directly addressing inference economics and making long-context workloads cost-competitive
• Represents meaningful efficiency inflection for open-weight models and signals architectural innovation beyond scale as competitive lever