24-hour AI news at a glance: ByteDoubao 1.8 and domestic AI entrepreneurship policies are accelerating, focusing on minor safety, watermark traceability and changes in the computing power supply chain overseas
In the past 24 hours (December 18 to December 19, 2025), domestic AI progress has focused on the upgrade of multi-modal capabilities of large models, the increase in entrepreneurship and computing power supply of local policies, and breakthroughs in new chip research; Overseas, leading model manufacturers have strengthened youth protection and content traceability, expanded AI infrastructure cooperation, and attracted attention around the policy trends of advanced chip sales to China.
- Byte Volcano Engine released the bean bag large model 1.8 and multi-modal generation series
Doubao Model 1.8 emphasizes stronger multimodal and agent capabilities, and improves long context and multi-step task processing. The simultaneous launch of image and video generation model series shows that domestic manufacturers are extending their "model capabilities" to the "usable productivity tool chain".
- Shanghai Xuhui released the "Super Entrepreneur Community" initiative to support the in-depth application of AI
The policy focuses on the supply of resources such as fee reductions, start-up funds and computing power vouchers, and supports talent services and events. For local AI entrepreneurship, the core benefit is to reduce the cost of early trial and error and accelerate the implementation from prototype to scenario.
- Shanghai Jiaotong University disclosed the research direction of all-optical large-scale intelligent generation chips
The research aims at higher energy efficiency and larger-scale generative computing needs, and responds to the contradiction between "computing power and energy consumption". If the subsequent engineering progresses smoothly, it may provide new hardware route imagination space for large model inference and training.
- Bairong Yunchuang released the enterprise-level AI Agent strategy and "result cloud" platform
The business narrative of enterprises shifting from "buying tools" to "buying results" is clearer, focusing on large-scale deployment in marketing, customer service, human resources, legal affairs and other positions. The key to whether this path can run through is to deliver quantifiable indicators and stable security compliance capabilities.
- OpenAI and Anthropic promote the identification and protection mechanism of minors
The two companies have strengthened their youth safety policies and identification methods for underage users, respectively, to reduce the harm caused by inappropriate advice and high-risk conversations. This move also reflects the increasing pressure of global regulators and public opinion on the "age hierarchical governance" of AI platforms.
- Google Gemini has added video watermark verification capabilities to strengthen SynthID traceability
Users can upload videos and check whether they contain platform watermark clues, and the system will mark the clips suspected of having watermarks. Content sourcing expands from "images" to "videos," helping to reduce deepfake dissemination costs and drive industry standardization discussions.
- Meta is reported to be developing a new generation of image and video models and a next-generation large language model
Meta was revealed to be promoting image and video generation models and text models with stronger coding capabilities at the same time, and exploring the direction of "world model". Multi-line parallel means that the competition of large manufacturers is shifting from a single-point model competition to a "multi-modal product matrix + developer ecosystem" competition.
- The U.S. Department of Energy promoted the Genesis Mission and signed AI cooperation with a number of technology companies
The collaboration covers cloud, chip, and cutting-edge model power, with the goal of using AI to accelerate research and serve energy and security issues. National scientific research demand is becoming an important landing field for cutting-edge models and computing power platforms.
- Mistral released Mistral OCR 3 to improve document understanding and structured restoration
The new version emphasizes high-fidelity extraction of document text and image content, and enhances structured output capabilities such as tables. Document intelligence is still one of the most "rigid" entrances to enterprise AI, and the evolution of OCR to "understandable document structure" is a trend.
- The United States launched a cross-departmental review of the sales of advanced NVIDIA H200 to China
The relevant review is regarded as a new variable that may appear in the advanced computing power supply chain policy, and the market and industry will pay attention to the scope and conditions of licensing. For the AI industry, the predictability of hardware supply will directly affect the cost curve of model training and inference.
Frequently Asked Questions (Q&A)
Q: What is the most noteworthy main line in the last 24 hours?
A: One main line is "content and youth security governance upgrade", and the other main line is "the uncertainty of computing power and hardware supply is rising", and the enterprise side accelerates the encapsulation of model capabilities into deliverable business results.
Q: What is the difference between domestic and foreign hot spots?
A: China is more inclined to "application implementation and policy/ecological support", while foreign countries are more focused on "platform responsibility, traceability governance and infrastructure cooperation", and superimpose the game of chips and export policies.
Q: What is the most direct enlightenment for enterprises to implement AI?
A: From a single model trial to "Agent process + quantifiable indicator delivery", data compliance, content security, and permission audit are incorporated into product design in advance.
Q: What should be the next priority for developers and entrepreneurs?
A: Priority is given to the construction of multi-modal capability orchestration, tool calling and evaluation systems, and the default capabilities for content traceability and security alignment are established to reduce the risk of rework caused by future compliance and platform rule changes.
