24-Hour AI News: DeepSeek's "Optical Compression" Breaks the Circle, Industry Edition Claude and AI Data Center Launch New Products

In the past 24 hours (October 21), China has seen multiple breakthroughs, including model open source, embodied intelligence, and energy large-scale model applications; Asia-Pacific and Europe and the United States have focused on AI data center cooperation, vertical industry model launches, platform-level function updates, and AI compliance reminders for election scenarios, showing an overall dual trend of "technology implementation + parallel governance."

1. DeepSeek Open Source DeepSeek-OCR: Proposing a New Paradigm of "Contextual Optical Compression"

1. Released and open-sourced on October 20, it adopts a two-stage architecture of DeepEncoder+DeepSeek3B-MoE-A570M.
2. When the compression ratio is less than 10×, the OCR decoding accuracy is about 97%; when it is increased to 20×, it is about 60%.
3. It was disclosed that a single A100-40G can generate over 200,000 pages of training data per day, verifying the feasibility of "replacing text tokens with pixels."

2. Yushu Technology releases a new generation of humanoid robot H2

1. H2 is approximately 180cm tall and weighs approximately 70kg. A bionic face and clothing demonstration has been added.
2. The 31 joints in the whole body improve flexibility compared with the previous generation, allowing you to complete continuous action demonstrations such as dance and martial arts.
3. Many media outlets followed up, and the industry focused on its cost-effectiveness and scalability in the implementation of embodied intelligence.

3. Beijing Humanoid Robot Innovation Center Open Sources Embodied World Model WoW

1. It was unveiled on October 21 and is positioned as the "brain" of the "Tiangong" humanoid robot.
2. Open source for the industry, helping robots quickly learn complex skills and physical causality.
3. Supporting the world model concept, emphasizing the application orientation of "from running to being easy to use".

4. Envision launches the world's first "AI wind and storage integrated machine"

1. Released on October 20 at the Beijing Wind Energy Exhibition, it integrates wind turbines, energy storage and the "Tianshu" energy model.
2. Officials claim that it can reduce the cost per kilowatt-hour, increase power generation and trading revenue.
3. It has the capability of “deep networking and isolated island operation” and is suitable for extreme working conditions and green electricity direct connection scenarios.

5. Beijing launches its first domestically produced AI classroom

1. On October 21, Beijing No. 20 Middle School was officially opened, and the class examples demonstrated AI-assisted traditional cultural teaching.
2. Focus on the "human-machine collaboration + domestic computing power" solution to serve multi-disciplinary teaching in middle schools.
3. It marks the transition of “AI in the classroom” from pilot to large-scale implementation.

6. Several AI-related national standards were updated on October 21st

1. The standard information platform shows that the update date of items such as "General Requirements for Trustworthy Embodied Intelligence, Embodied Intelligence Evaluation Guidelines, Embodied Intelligence Application Framework and Interface" is October 21.
2. At the same time, educational items such as "Technical requirements for basic education large-scale model systems" appear.
3. It shows the simultaneous advancement of "standard traction + scenario implementation", covering embodiment, education and industry evaluation.

7. Hitachi and OpenAI announce strategic partnership: Focus on global AI data center expansion

1. Company news was released on October 21, positioning the collaboration of energy, digital intelligence and data center.
2. Focus on global layout and high-performance computing infrastructure.
3. Release the integrated investment signal of “computing power-energy-data center”.

8. Anthropic Launches Claude for Life Sciences and Expands Claude Code to Web and Mobile

1. The industry version of Claude is aimed at life sciences and supports reading literature, generating regulatory documents, and connecting experiments with genetic tool chains.
2. It is reported that some pharmaceutical companies have conducted internal pilot projects to significantly shorten the document processing cycle (reported).
3. Claude Code has added a new web and mobile "research preview" for Pro/Max users.

9. Microsoft's 2025 Digital Defense Report focuses on the rising threat of AI

1. The report states that adversaries and criminal groups are accelerating the use of AI to manipulate public opinion and conduct cyberattacks.
2. In July alone, we identified over 200 cases of foreign entities using AI to create false content, a significant increase both year-on-year and month-on-month.
3. Simultaneously advocate that the government introduce mandatory safety standards and compliance requirements in the procurement of AI tools.

10. Google launches Gemini "multi-sheet AI analysis" capability for Sheets

1. Multiple tables can be selected and analyzed in a linked manner on the same page, and formula generation, charts, and cross-table editing are supported.
2. Gradually push it to commercial and educational subscription groups.
3. Cooperate with the Workspace family's "AI Assistant" to enhance the intelligence of office data.

11. The Dutch data regulator warned before the election: Don't rely on AI chatbots to make voting decisions

1. On October 21, a reminder was issued that chatbots were "unreliable and unverifiable" as voting aids.
2. Tests show that output is biased in multiple scenarios.
3. Add another case study to the European “AI × election” compliance and risk governance.

12. Samsung will release Project Moohan XR headset at a global event tonight, emphasizing multimodal AI

1. The event is scheduled for the evening of October 21 (Eastern Time), with the theme of "Worlds Wide Open."
2. The official preheating points to "AI-driven immersive intelligence" and Android XR ecological collaboration.
3. The industry is focusing on its system-level integration with mobile AI, 3D capture and other capabilities.

Frequently Asked Questions (Q&A)

Q: What are the technical highlights of DeepSeek-OCR? Does it save more tokens?

A: The core is to render long text into images and then carry information with visual tokens. The measured OCR accuracy is about 97% when the compression rate is less than 10×, which significantly reduces the token occupancy of LLM long context. A single A100-40G card can generate up to 200,000 pages of data per day.

Q: What specifically can Anthropic’s Claude for life sciences do?

A: It aggregates scientific research ecosystems such as Benchling, 10x Genomics, and PubMed to support solution selection, data table analysis, compliance document generation, and process automation. The goal is to enable a larger proportion of life science R&D work to be completed on Claude (reporting caliber).

Q: What does Microsoft's latest cyber threat assessment mean for organizations?

A: As threats characterized by "AI-generated content + automated attack chains" become normalized, organizations need to incorporate model security, data integrity, and supply chain protection into their baselines, and implement "security requirements + third-party evaluations" in the procurement phase.

Q: Which teams are suitable for Google Sheets’ “Multi-sheet AI Analysis”?

A: Suitable for finance, operations, and teaching teams who need to connect and analyze data across multiple tables. AI can generate cross-table formulas and charts based on selected areas, reducing manual processing and interpretation costs.

Q: Will the Dutch regulatory warning affect the use of AI in other election scenarios?

A: In the short term, it will push the platform to strengthen disclaimers and safety net mechanisms during the election season, and in the long term it may promote the transparency, verifiability and neutrality constraints of "AI voting tools."