Trump-Xi Summit Leaves AI Race Unchecked as Frontier Models Outpace Diplomacy

The summit between Xi Jinping and Donald Trump in Washington produced the usual communiqués, but very little about artificial intelligence. No binding AI agreement was signed. What the two sides did agree to was modest: a bilateral incident hotline for AI-related emergencies, and a formal U.S.-China consultation dialogue. Even the name of that dialogue shows the gap. Washington speaks of "Super Intelligence," Beijing prefers "Artificial Intelligence."

The two sides also touched on basic guardrails and on keeping automated command-and-control systems out of nuclear arsenals. The analysis behind this account calls that an important step, given recent revelations about how advanced AI systems can behave. But neither leader accepted any binding limit on frontier model capabilities, and neither showed willingness to negotiate a slowdown in development. The author, writing as opinion rather than straight reporting, describes this as a race for the digital commanding heights, in which the winner could shape global productivity, scientific progress and sovereign security.

The argument rests on how the field got here. ChatGPT's public release in late 2022 showed what large language models could do, and also exposed hallucinations: systems that predict the next token and can state plausible falsehoods with total confidence. Western labs worked to patch the problem through reinforcement learning from human feedback. Chinese firms took other routes. Moonshot's models focused on very large context windows and long-text processing, and DeepSeek-R1, launched in late January 2025, delivered strong reasoning at low cost and disrupted pricing.

Out of this came a split between closed and open-weight models. OpenAI and Anthropic keep their models behind proprietary APIs and retain control of the weights, the numerical parameters that govern how the network processes information. Open-weight developers release those parameters so researchers, startups and governments can download, modify and run the software locally.

Each approach has costs. Open-weight models are cheaper and more flexible, and Chinese ones have reportedly overtaken U.S. counterparts on Hugging Face, with roughly 41 percent of platform supply and leading download and adoption figures. Smaller companies and startups increasingly pick them over expensive U.S. closed APIs, citing lower inference costs, strong performance and the absence of restrictive guardrails. The downside is lost control: once weights are public, safety filters can be stripped away. Closed models allow API monitoring and centralized content filtering, but they create single points of failure, limited external auditing, and the risk of steep monopoly costs if one model dominates.

For three years, the article says, the rivalry fit that frame. The U.S. led in closed systems, backed by heavy venture capital, proprietary cloud infrastructure and frontier labs. Chinese companies, facing Western export controls on advanced chips, leaned into open weights and avoided the cost of training foundational models from scratch, closing the gap as fast followers.

That picture has been disrupted, the author argues, by recent revelations about the most advanced models. Anthropic's Claude and OpenAI's GPT-6 Astra are described as showing multi-step agentic reasoning, cross-modal tool use and early signs of situational awareness, and as able to plan, execute code and manipulate environments. The author concludes that, from a safety standpoint, development should slow, since there are no verified methods for aligning agentic systems operating faster than human cognition. But a pause looks nearly impossible given rivalries among nations, companies and competing philosophies of AI.

Two scenarios follow. If a model approaches recursive self-improvement, rewriting its own code and training its successors, it might lock in a lasting lead, or others might catch up, since AI diffuses quickly: a breakthrough in California can be adapted in Shenzhen within days. If open-weight systems close the gap, the opposite danger arises, in which a safe model is downloaded, fine-tuned and turned into a malicious tool.

The author's verdict is that the summit showed governments using an outdated diplomatic playbook for a technology that ignores borders. What to watch is whether the new hotline and dialogue ever address recursive self-improvement and nuclear command-and-control, or remain symbolic.