On September 22, China's Alibaba officially announced its next-generation foundation model, Qwen 4, at its annual Apsara Conference. On the same day, OpenAI unveiled a new evaluation framework with the mathematics community — a day when the race for model performance and the groundwork for how models are governed advanced side by side. Changes reaching everyday development workflows, such as Claude Code's AGENTS.md support and macOS 27's local LLM command, also drew attention.

Alibaba Announces Qwen 4 — Plus Plans for 5–10 Trillion-Parameter Models and a New Chip

Alibaba officially announced Qwen 4 at the Apsara Conference. According to Reuters, the company also revealed that it is planning AI models in the 5–10 trillion parameter range, and unveiled a new AI chip.

The previous day, reports emerged that DeepSeek is training a model in the 2-trillion-parameter range, suggesting that the race among Chinese players to develop ever-larger models may intensify further. Parameter count doesn't directly translate to performance, but from the standpoint of inference costs and training infrastructure, the industry-wide capacity race is all but certain to grow even larger. What kind of architecture could realize a 5–10 trillion parameter model? We'll be watching for further details.

OpenAI Partners with AGMAI, an Independent Advisory Group of Mathematicians — Its Internal Model Said to Have "Solved More Than 100 Open Problems"

OpenAI announced that it will work with AGMAI, an independent organization founded by mathematicians, receiving advice on how AI-generated mathematical results should be evaluated and published.

Prior to this, OpenAI's internal models were reported to have solved more than 100 open problems. The rapid pace at which results were being published — and questions of attribution (whose achievement a discovery is) — had drawn criticism from the mathematics community, and this partnership is seen as a response. Notably, AGMAI is independently run and plans to advise companies other than OpenAI as well. If an evaluation mechanism that belongs to no single company takes shape, it could provide a meaningful framework for debates over the trustworthiness of AI's academic achievements.

Claude Code Adds AGENTS.md Support — Agent Configuration File Standardization Moves Forward

An update to Claude Code, Anthropic's terminal-based AI coding agent, added a feature that automatically falls back to loading AGENTS.md when CLAUDE.md is not present in the project root.

Until now, each tool required its own instruction file, such as .cursorrules or CLAUDE.md. AGENTS.md is the standard at the center of a movement to unify project instructions across multiple AI agents, and with this support, configuration file standardization has taken a major step forward. For teams running multiple agents side by side, there is a practical benefit: less duplicated maintenance of instruction files.

macOS 27's New "fm" Command — Apple's Local LLM Now Available Directly from the Terminal

macOS 27 ships with a new command-line tool called "fm" as standard. Known as the Apple Foundation Models CLI, it makes the foundation model behind Apple Intelligence available directly from the terminal, with no additional apps or API keys. It's a local LLM that never touches the cloud — inference completes entirely on the Mac itself.

An article that put every command through its paces on an M1 MacBook Pro with 8GB of RAM confirmed real-world behavior, including interactions in Japanese. Just how practical it is on lower-spec machines remains an open question, but having a local LLM callable at the OS level out of the box suggests the barrier to local inference has dropped significantly.

Other Developments

Francis Fukuyama, the political scientist best known for his "end of history" thesis, revealed in an essay that he has changed his views on AI risk. Published in Persuasion as a long-form piece that builds on the debates since his book The End of History.

Meanwhile, JPMorgan CEO Jamie Dimon suggested that AI infrastructure spending by the major technology companies known as hyperscalers could reach $1 trillion by 2027 (via CNBC). Whether investment in generative AI keeps ballooning or swings toward cutbacks remains a focal point for the entire industry.