24-hour AI news: Spiritual Breath SDG large model released, OpenAI released hallucination research

24-hour AI news at a glance: domestic models and industrial exhibitions are new at the same frequency, overseas supervision and research progress are carried out at the same time.

In the past 24 hours, China has released a large model serving the United Nations Sustainable Development Goals, and the Industrial Internet and the Smart Expo will focus on "artificial intelligence +"; Overseas, the state attorney general put pressure on OpenAI on the safety of minors, and OpenAI released research on the mechanism of "hallucinations" and was exposed to large-scale capital expenditure plans.

1. China released the "Lingxi Sustainable Development Goals Model"

1. It was officially released at FBAS 2025 in Beijing on September 6, positioning itself as a comprehensive monitoring and decision support for the United Nations Sustainable Development Goals (SDGs).
2. Technical features: multi-modal data fusion, retrieval augmented generation, spatio-temporal analysis and causal inference, built-in literature review, index analysis, report writing, three-dimensional interaction, and intelligent narrative 5 agents.
3. Significance: China has formed a general scientific research and policy evaluation base in "AI + sustainable development", and has been implemented for global application scenarios.

2. The 2025 Global Industrial Internet Conference opened in Shenyang, with the theme of "Artificial Intelligence +"

1. It will be held from September 6 to September 8, and the exhibition area will focus on industrial applications such as large models, AR/VR, and embodied intelligent robots.
2. Emphasize the connection between supply and demand and the transformation of achievements, focusing on new industrialization and the integration of data and reality.
3. For the industrial chain: promote the AI of industrial scenarios and the collaborative iteration of equipment and platforms.

3. China Unicom appeared at the World Intelligent Industry Expo and released AI infrastructure and industry solutions

1. The Chongqing Smart Expo opened on September 5, and operators exhibited AI computing power, platforms and industry application solutions.
2. Focusing on "AI infrastructure-AI technology-AI application", emphasize the co-construction of ecology with industrial partners.
3. On-site trend: AI large models "get on the car", and on-board capabilities such as intelligent interaction and emotional perception have become highlights.

4. Shandong released generative AI service registration information: 4 new models, Jinan ranked first

1. As of September 2, the province has completed 4 new provincial registrations; Jinan has 2 models.
2. Jinan has completed the filing of a total of 11 items with the Cyberspace Administration of China, ranking first in the province in terms of total volume and increment.
3. Key points of the policy: Applications that call the filed model through API can be "registered and managed", and the number must be indicated in a prominent position when launched.

5. The industrialization of "embodied intelligence" has become the focus: the Science and Technology Innovators Conference was held in Beijing

1. It was held on September 5 with the theme of "Embodied Intelligence, a New Engine for Industrial Intelligence".
2. Signed a contract to establish the "China-Italy Robot (Beijing) Research and Innovation Center" to promote international cooperation.
3. The topic focuses on scientific research transformation, industrial landing and investment and financing docking.

6. OpenAI released research: why large models "hallucinate"

1. A public paper published on September 5 pointed out that the existing training and evaluation mechanism rewards "guessing" rather than "restraint when uncertain", inducing hallucinations.
2. Suggest reshaping the evaluation: high penalties for "self-confidence mistakes" and partial points for reasonable "abandonment/indication of uncertainty".
3. Conclusion: The hallucination rate can be significantly reduced through calibration and evaluation transformation, especially in complex real tasks.

7. OpenAI reportedly plans to spend about $115 billion in cumulative cash by 2029

1. The media said that the expenditure will exceed $8 billion in 2025, and plans to develop self-developed acceleration chips and self-built/cooperative data centers.
2. Cooperation context: Promote computing power and data center projects with partners such as Broadcom, Oracle, and Google Cloud, involving 4.5GW facilities and larger-scale "Stargate" initiatives.
3. Industry implications: Computing capital expenditure has accelerated its spillover to semiconductors, energy and real estate infrastructure.

8. California and Delaware attorneys general send a letter to OpenAI: The safety of minors must be strengthened

1. On September 5, the attorneys general of the two states sent a letter to OpenAI's board of directors after the meeting, requesting improved child and adolescent protection measures.
2. Background: Previously, the attorneys general of 44 states jointly issued an industry warning about inappropriate interaction with chatbots.
3. Regulatory trend: During the review of OpenAI's governance and capital structure adjustment, security compliance was placed as a "precondition".

9. British media industry: Google search "AI" impacts traffic

1. On September 6, it was reported that AI summaries/AI models reduced backlink clicks, and organic traffic in some media fell significantly.
2. Industry response: litigation, regulatory complaints and licensing cooperation go hand in hand, and some media are exploring self-built AI Q&A and licensing models.
3. Structural proposition: The AI of search entrances reshapes the content distribution and advertising model, and the risk of "Google Zero" in the industry intensifies.

10. Oracle earnings outlook: analysts are optimistic about AI cloud growth

1. The financial report will be disclosed on September 9; The year-to-date return is about +37%, and many institutions maintain a "buy".
2. Highlights: GPU cluster cost-effectiveness, full-stack AI capabilities and locked large orders; High capital expenditure suppresses profits in the short term.
3. Risks and opportunities: The rhythm of orders and gross profit fluctuations coexist, and the long-term compound growth rate is expected to rise.

Frequently Asked Questions (Q&A)

Q: What are the different designs of the Lingxi SDG large model from general-purpose LLMs?

A: The integration of spatio-temporal causality and multimodality for the SDGs emphasizes indicator systems, policy evaluations, and three-dimensional interactive narratives, with built-in 5 types of agents to support the closed loop of scientific research and governance (released at the Beijing conference on September 6).

Q: What are the direct implications of the state attorney general's letter to OpenAI?

A: In the short term, we will promote the upgrading of minors' safety functions, parental control and crisis response; In the medium term, "safety accountability" is taken as a hard constraint in the approval of organizational governance/capital operation.

Q: How has OpenAI's "hallucinations" research changed the evaluation of the model?

A: The core is to explicitly explicitly formulate the cost of "confidence error" and give the "reasonable uncertainty" part to encourage the model to answer less when uncertain rather than guessing, thereby reducing the hallucination rate.

Q: Why is the capital market interested in Oracle's AI cloud?

A: Its GPU cluster and integrated stack bring price/performance advantages, and there are large contract clues, but the high-intensity CapEx short-term gross profit will be verified, and the financial report will verify the order fulfillment and profit curve.