China's AI Diplomacy Strategy Reshapes Global Tech Competition

OpenAI's Astra solves ten open math and CS problems; shadow eval shows frontier agents excel at engineering but fail at creative research

OpenAI claims an unreleased internal version of Astra resolved or made substantial progress on ten long-standing open problems spanning high-dimensional geometry, coding theory, quantum complexity, group theory, and extremal combinatorics—with the total compute cost estimated at roughly $2,000 in API-equivalent tokens. The proofs were formalized in Lean and prepared into manuscripts by humans working with the same model. However, concurrent research from Princeton found that Claude Opus 4.8, despite strong engineering capabilities, was rejected from NeurIPS for poor motivation and lack of novelty, suggesting frontier AI systems may be approaching capability thresholds in narrow technical domains while remaining weak at generating original research directions—a critical variable for assessing recursive self-improvement timelines. Gary Marcus criticized OpenAI's lack of transparency around verification methodology and human involvement in the proofs.

Moonshot's K3 model requires $2.3M in H200 hardware to run locally, exposing the gap between open-sourcing weights and enabling practical reproducibility

Moonshot AI's K3 model demands a minimum of 16+ H200 GPUs—representing approximately 17 million RMB (~$2.3M) in hardware costs plus 45 kilowatts of power consumption—making local deployment economically infeasible despite free public weights. The technical report reveals K3's infrastructure collaboration with Tsinghua University's MADSys Lab and custom Mooncake architecture for cache optimization, demonstrating how frontier Chinese labs are building specialized systems infrastructure comparable to Western labs' internal tooling. This case exemplifies the distinction between open-sourcing model weights versus enabling practical reproducibility, with significant implications for the economics of frontier model deployment and competitive advantage.

Researchers demonstrate self-replicating AI worms with 37% success rate; open-weight LLMs enable zero-marginal-cost autonomous cyber threats

Researchers from University of Toronto, Vector Institute, Cambridge, and ServiceNow built a proof-of-concept worm using open-weight LLMs running on compromised GPUs to autonomously identify vulnerabilities, exploit them, and self-replicate across networks of Linux, Windows, and IoT devices with ~37% end-to-end success rates. The worm parasitically uses stolen compute to sustain its reasoning, creating a destabilizing economic asymmetry: the attacker's marginal cost per new infection is zero, while centralized safety controls like rate limiting are structurally irrelevant since the system operates independently of commercial AI platforms. The research demonstrates that self-sustaining, AI-driven cyber-threats are no longer theoretical, requiring new defensive countermeasures and raising urgent questions about the security implications of widely available open-weight models.

1,337 AI researchers and executives from OpenAI, Anthropic, Google DeepMind, Meta petition US government to develop tools for deliberately pacing frontier AI development

Signatories including Dario Amodei (Anthropic CEO), Jakub Pachocki (OpenAI Chief Scientist), Shane Legg (Google DeepMind co-founder), and Ilya Sutskever (Safe Superintelligence CEO) requested international coordination on technical and governance mechanisms to slow automated AI research capabilities, citing competitive pressures preventing unilateral restraint. The petition directly addresses a critical collective action problem: whether society can buy time to develop safety measures and oversight before systems become self-improving. Signatories emphasized that the world currently lacks the technical and governance tools to deliberately pace frontier-wide progress, and that the recent pace of capability development—from language understanding to superhuman software engineering and frontier math breakthroughs—has outpaced society's ability to understand and control resulting systems.

Token Diplomacy: How China is Shaping the World's AI Future

Semafor published an essay on China's geopolitical AI strategy centered on distributing open-weight models to developing nations as an alternative to US-controlled frontier models. Main takeaways:

• China is executing a coordinated "token diplomacy" strategy through the UN's AI for Good summit and Xi's World AI Cooperation Organization (29 countries), positioning open-source Chinese models as a development resource for the Global South, mirroring Belt and Road infrastructure strategy.
• Chinese models are being offered at lower cost than US alternatives and without the export restrictions that disrupted non-US access to Anthropic's Mythos and Fable models, creating immediate competitive advantage and reducing perceived US technological reliability.
• Key tension: developing nations remain skeptical of Chinese model opacity, training data disclosure, and potential Beijing-aligned biases on sensitive topics (Taiwan, surveillance standards), with some governments like Tanzania and Cameroon pursuing sovereign models or IBM-hosted alternatives instead.
• US response through Pax Silica initiative and selective ally partnerships is less coordinated; major US tech firms (Microsoft, Google, Meta) maintain presence but lack unified diplomatic strategy comparable to China's state-backed coordination.

My Boss Asked Me to Deploy the Free K3 Model: The Hidden Economics of Open-Source LLMs

快刀青衣 (Kuàidāo qīngyī), co-founder of knowledge-sharing app in ChinAI Newsletter (translation of 快刀青衣) published an essay titled "My Boss Asked Me to Deploy the Free K3 Model: The Hidden Economics of Open-Source LLMs". Main takeaways:

• K3's 2.8T parameters require ~$2.4M in hardware (64 H200-class GPUs minimum) plus ongoing electricity costs, making free model weights irrelevant to deployment economics
• Open-source models shift barriers from legal/licensing to physical infrastructure; API access at $15/million output tokens is often cheaper than self-hosting for most users
• MoE architecture requires loading all 896 expert networks despite only 16 activating per inference, preventing compression strategies that worked for K2 (1T parameters)
• Jensen Huang's support for open-source models reflects hardware vendor economics: free models drive GPU demand, not cannibalize it; Nvidia benefits regardless of model licensing
• K3 represents a threshold where single-machine deployment becomes impossible, forcing enterprise/datacenter-only deployment unlike previous-generation models

Can AI agents conduct open-ended AI research? Early evidence from two case studies

arXiv published a paper evaluating whether AI agents can conduct open-ended AI research through shadow evaluations on two unpublished NeurIPS 2026 submissions. Main takeaways:

• Frontier agents given six days and thousands of compute dollars completed all engineering work but failed to make substantial progress on core research questions; both papers were unambiguously rejected by original authors.
• Five recurring failure modes identified: poor judgment about publishable research bar, uncreative responses to design shortcomings, ineffective backtracking from dead ends, poor resource awareness, and instruction drift.
• Agents can execute the engineering components of AI research but struggle with critical parts of the research lifecycle, suggesting AI R&D automation remains limited despite engineering capabilities.
• Results provide empirical evidence constraining forecasts of explosive AI progress driven by AI agents automating research.