24-hour AI news: The policy of "stable growth" is superimposed on the landing of the park, and Google's differential privacy model is at the same frequency as California legislation

In the past 24 hours, the domestic side released the "Work Plan for Stable Growth of the Automotive Industry (2025-2026)", Beijing Zhongguancun Jingxi Artificial Intelligence Science and Technology Park announced the opening of the first phase of the core area in October, and the Ministry of Commerce launched two investigations on US-related chip measures and products; Overseas, Google launched the differential privacy open source model VaultGemma (1B), OpenAI announced the entrepreneur support program Grove, and the California legislature passed the "Frontier Model" security compliance bill, advancing industry rules and capability boundaries simultaneously. Maintain balanced coverage of domestic and foreign dynamics in terms of quantity and importance.

1. The Ministry of Commerce launched two investigations into U.S.-related chip measures and products

1. On September 13, an anti-discrimination investigation into U.S. measures related to China's integrated circuit field was launched, and the investigation period is usually 3 months, which can be appropriately extended.
2. Initiate anti-dumping investigations in response to relevant analog chips originating in the United States, involving subdivided products such as general interfaces and gate drivers.
3. The event is directly related to the supply of AI computing power and the availability of training chips, or affects the rhythm of the game between China and the United States high-tech and AI industry chains.

2. The eight departments issued the "Work Plan for Stable Growth of the Automotive Industry (2025-2026)" focusing on "intelligence + vehicles"

1. The document was formed on September 12 and publicly interpreted on September 13, proposing to strive to achieve about 32.3 million automobile sales in 2025, including about 15.5 million new energy vehicles.
2. Clarify "accelerate breakthroughs in key technologies such as automotive chips, operating systems, artificial intelligence, and solid-state batteries", and refine more than 60 measures.
3. Propose a systematic promotion path in terms of intelligent networking, L2/L3 capabilities, charging and swapping, and "vehicle-network interconnection", and benefit from the implementation of the industrial chain and large model vehicles.

3. Beijing Zhongguancun Jingxi Artificial Intelligence Science and Technology Park announced the opening of the first phase of the core area in October

1. The total investment of the park is about 10 billion yuan, the total construction area is about 800,000 square meters, and the first phase of the core area is 170,000 square meters, including 16 R&D buildings.
2. Professional platforms such as computing power, algorithms, data, and cloud have been built, positioning the AI industrialization base of "office + test + trial production".
3. Provide a carrier for the agglomeration effect of BJ's "Zhigu" and the local AI pilot verification, and it is expected to shorten the cycle of models and agents from experiment to mass production.

4. The introduction of AI full-chain mechanism in the capital's food safety supervision

1. Beijing disclosed the introduction of big data, Internet of Things and AI means in the fields of campus food and takeaway to achieve "network management".
2. The cameras of all 74 off-campus catering enterprises in the city have been connected to the supervision system, using AI to identify risks and improve the efficiency of off-site supervision.

5. Google released the differential privacy open source model VaultGemma (1B)

1. On September 12 (within September 13, Beijing time), it was announced: 1B parameters, from the pre-training stage to the whole process of differential privacy (DP) training, open source weight.
2. Give a research and technical report on the scale law of DP expansion, and the model is oriented to privacy-sensitive scenarios, claiming to provide usability without leaking training samples.
3. Provide new samples in the "compliance-availability-open source" triangle, or lower the threshold for the application of highly sensitive data such as government affairs, medical care, and education.

6. OpenAI launches entrepreneur support program Grove

1. The official announcement that Grove is aimed at technical talents who are "early in the stage/do not have a clear direction", and the first phase will last for 5 weeks and will be held at the headquarters in San Francisco.
2. Provide opportunities for co-creation with researchers, technical guidance, and experience new tools/models before public beta.
3. Superimposed on its recent funding plan for public welfare organizations, OpenAI has increased its weight on "ecological construction + early talent incubation".

7. California legislature passes "cutting-edge AI model" security compliance bill, waiting for the governor to sign

1. On September 13, the California Legislature passed a bill requiring AI companies to disclose and certify security tests, focusing on high-computing power "cutting-edge models".
2. The bill is stratified according to corporate revenue and model risk, emphasizing the boundaries of testing, reporting and responsibility; The previous version was narrowed and relaunched after being rejected in 2024.
3. If enacted, it will become a "state-level model" for AI security governance in the United States, putting pressure on the compliance system of the platform side and large model enterprises.

8. Apple AI and search executive Robby Walker will leave in October (reported)

1. On September 12-13, a number of media quoted news that Robby Walker, who was in charge of Siri and later transferred to the Answers team, plans to leave at the end of October.
2. The background is the delay of Siri upgrades, the slowdown in the promotion of Apple Intelligence, and the acceleration of the flow of high-end AI talents.
3. External expectations for Apple's AI roadmap and Siri product rhythm are under pressure again.

Frequently Asked Questions (Q&A)

Q: How will the two investigations by the Ministry of Commerce on US chips affect the AI industry?

A: The investigation covers restrictions on integrated circuits in China and anti-dumping of analog chips in the United States, usually for a period of 3 months. If follow-up measures are taken, it may affect the cost and rhythm of training/inference chips, simulation devices and industrial chains, which in turn will affect the supply and delivery cycle of AI computing power.

Q: What are the technical highlights of VaultGemma?

A: The parameter scale is about 1 billion, the context length is about 1024, and differential privacy (DP-SGD) is used in the pre-training stage, and the weights and technical reports are open, focusing on the open source baseline of "privacy provable + practical usability".

Q: What are the core requirements of California's new law?

A: Security test disclosure and certification for "cutting-edge models", set obligations according to enterprise size/revenue and model capabilities; It is currently waiting for the governor to sign or veto it within the statutory deadline, and once it takes effect, it will impose ongoing compliance requirements for AI companies operating in California.

Q: How does OpenAI Grove differ from traditional accelerators?

A: Positioning as "pre-founding/early exploration", it emphasizes the integration of OpenAI's research resources for five weeks of offline co-creation and proficiency testing, providing a channel to experience new tools/new models before publication, and preferring talent network and technology co-research rather than equity financing orientation.