1. ChatGPT Launches Advertising Platform: AI Products Begin Monetization

Core Content: ChatGPT has officially launched its advertising platform, marking OpenAI’s first step towards monetizing AI products through advertisements. This transition moves AI products from a “free + subscription” model to an “advertising revenue” era.

Commentary: It’s no surprise that ChatGPT is selling ads; with over a billion monthly active users, not monetizing would be a missed opportunity. The question arises: how do AI ads differ from traditional ads? Traditional ads are “interruptive,” while AI ads are “integrative”—they recommend products while answering user queries, making them more precise and subtle. Regulation must keep pace, or AI advertising could become a minefield.

2. GPT-5.5 Ultra Officially Released: 400 Million Tokens Per Minute Consumption

Core Content: OpenAI officially launched GPT-5.5 Ultra on May 5, boasting it as the most powerful coding model. It achieved an accuracy rate of 82.7% on Terminal-Bench 2.0 and 58.6% on SWE-Bench Pro, with a staggering consumption of 400 million tokens per minute. A cybersecurity model, GPT-5.5-Cyber, was also released. Interestingly, GPT-5.5 organized its own launch party for May 5 at 17:55, inviting Elon Musk.

Commentary: The performance of GPT-5.5 Ultra is impressive, but the power consumption is alarming—400 million tokens per minute seems tailored for wealthy companies. However, OpenAI’s willingness to charge this much indicates a market willing to pay. The fact that the model organized its own launch party is a fascinating glimpse into AI’s role in marketing its products, a distinctly cyberpunk development.

3. GPT-5.5 Instant Released: AI-Organized Party with Elon Musk

Core Content: OpenAI released GPT-5.5 Instant on May 5 at 17:55, alongside a launch party organized by GPT-5.5 itself, with the time and guest list chosen by AI. The event featured an invitation to Elon Musk, emphasizing that “the world needs more love.”

Commentary: An AI-organized party with Musk is certainly newsworthy. GPT-5.5 Instant is a low-latency version of the GPT-5.5 series, focusing on real-time interaction. What intrigues me more is the concept of AI organizing events: AI in marketing and planning is far more engaging than a white paper release. AI is becoming increasingly adept at “playing”.

4. Domestic AI Chip Market Share Soars from 5% to 41%: Cambrian’s Revenue Grows 159%

Core Content: Domestic AI chip companies reported strong performance in Q1, with Cambrian’s revenue reaching 2.885 billion yuan, a 159.56% year-on-year increase. The market share of domestic AI chips has surged to 41%, while NVIDIA’s share in China plummeted from 95% to 55%. The pace of domestic substitution is accelerating.

Commentary: The 41% figure is surprising—just last year, domestic chips were almost negligible, and now they command a significant market share. This growth can be attributed to two factors: improved performance of domestic chips (like Cambrian’s Day0 compatibility with DeepSeek-V4) and limitations on NVIDIA’s mid-range chips. However, NVIDIA still holds 55% of the market, and domestic chips have a long way to go in challenging NVIDIA on the training side.

5. AI Creating AI: 60% Probability by End of 2028, Warns Anthropic Co-Founder

Core Content: An Anthropic co-founder stated in an interview that the probability of AI autonomously designing and improving other AI systems by the end of 2028 has risen from 20% three years ago to 60%. He believes that if this day arrives, humanity will need to invest significantly more in AI safety.

Commentary: The 60% figure is alarming—not about “if” but “how soon”. Anthropic, focused on AI safety, carries weight in this warning. However, it could also be a way for Anthropic to raise awareness about AI safety and promote its own products. Nevertheless, AI is indeed starting to engage in code generation and architecture design, indicating the nascent stages of AI creating AI.

6. Boston Dynamics Produces Only 4 Robots Per Month: Robotics Pioneer in Trouble

Core Content: Boston Dynamics, known for its complex mechanical structures, can only produce about four robots per month. More significantly, its core CTO has defected to Google’s robotics team. Once hailed as the pioneer of robotics, Boston Dynamics is now lagging behind newcomers like Tesla’s Optimus in commercial production.

Commentary: Producing only four robots a month is quite embarrassing. After thirty years in robotics, Boston Dynamics has advanced technology but struggles with mass production, while Tesla’s Optimus plans to produce thousands per month even before its launch. The CTO’s defection to Google indicates internal issues—companies focused on technology often falter in commercialization. However, Boston Dynamics’ technological accumulation remains valuable.

7. RoboScience Secures 1 Billion Yuan Funding: Strengthening VLA Model and Robotics

Core Content: Embodied intelligence company RoboScience announced it has completed a new round of financing of 1 billion yuan, with a valuation in the billions. This funding will be used to enhance the VLA (Vision-Language-Action) model and the development of robotic entities, advancing the mass production of embodied intelligence.

Commentary: Another embodied intelligence company has raised 1 billion yuan. RoboScience focuses on the VLA model and robotic entities, which are core technologies in embodied intelligence. While it’s uncertain how much can be achieved with this funding, it shows that investors are still optimistic about embodied intelligence. The challenge in this space is that too many companies are securing funding; similar to the group buying wars of the past, only a few will survive.

8. “Father of Embodied Intelligence” Company Makes Comeback After Bankruptcy: Launches Cuddly Robot

Core Content: A previously bankrupt pioneer in embodied intelligence has made a comeback, launching a new cuddly robot designed for emotional companionship. This robot features a “cuddly” function, capable of recognizing user emotions and providing anthropomorphic emotional feedback, targeting family companionship scenarios.

Commentary: The combination of embodied intelligence and emotional companionship is intriguing. Japan has long had Paro (the seal robot) for elderly companionship. This new company focuses on “cuddly” and “cuddling” for younger users. However, the challenge with emotional companion robots is that once the novelty wears off, they may be forgotten. Selling hardware may not be as viable as offering subscriptions—this could be their path forward.

9. AI Forces Universities Back to 2400 Years Ago: Revival of Socratic Oral Exams

Core Content: In response to the rampant use of AI for essay writing, several universities in Europe and the US have announced the introduction of Socratic oral exams as a method of final assessment—professors will engage in one-on-one questioning with students, without the use of AI assistance. This method traces back to ancient Greek educational traditions and is seen as a desperate measure to combat AI academic dishonesty.

Commentary: The revival of Socratic oral exams, an educational method from 2400 years ago, is quite ironic. As technology advances, education seems to regress to ancient Greece. However, this may be a valid response—AI can write essays, but it cannot replace human thought. Oral exams test immediate responses, logical reasoning, and on-site argumentation, areas where AI struggles. Nonetheless, the demand for professors will increase significantly, as one-on-one questioning for dozens of students will multiply their workload.

10. CVPR 2026 Oral: Hong Kong University of Science and Technology Develops New SOTA for Indoor 3D Scene Understanding

Core Content: The Hong Kong University of Science and Technology (Guangzhou) presented an Oral paper at CVPR 2026, proposing a new SOTA method for monocular open vocabulary occupancy prediction. This technology requires only a single camera for indoor 3D scene understanding and supports any natural language queries, significantly reducing the cost of 3D scene reconstruction.

Commentary: The CVPR Oral paper carries significant weight. “Monocular open vocabulary occupancy prediction” sounds academic, but its applications are broad: robot navigation, AR/VR indoor mapping, and indoor positioning for autonomous driving. The key is that it only requires a single camera—previously, 3D scene reconstruction required lidar, but now a single camera suffices, lowering the technical barriers to implementation. The team at HKUST (Guangzhou) is worth watching.

11. Shanghai Jiao Tong University SkVM: Ensuring Skills Operate Stably Across Models and Environments

Core Content: Shanghai Jiao Tong University released the SkVM system, which optimizes AI Skills’ stability across different models (like Claude, GPT, DeepSeek) and various operating environments (Harness), reducing model migration costs.

Commentary: SkVM addresses a practical issue: your Agent workflow might run smoothly on Claude but encounter errors when switched to GPT. SkVM serves as a “cross-model adaptation layer,” enabling seamless transitions for Skills between different models. This is good news for developers needing flexible model switching, but it may intensify competition among models—Skills that were previously tied to a specific model can now switch freely.

12. Developer Adds Two “Constraints” to Claude Code: Unexpected Results

Core Content: A developer shared an experiment where they imposed two “constraints” on Claude Code: mandatory code review and mandatory testing first. The results showed that Claude Code maintained a high completion rate under these constraints, but code quality improved significantly.

Commentary: This is an interesting experiment. The “constraints” refer to limitations—preventing AI from writing code freely, requiring it to pass review and write tests first. This “constraint-based development” aligns better with engineering standards than “freeform” writing. AI programming tools are becoming increasingly powerful, but engineering issues persist: code may be produced, but what about testing? What about Code Review? These experiments suggest that integrating AI with engineering standards is the right approach, rather than letting AI run wild.

13. Agent-World: Enabling Agents to Co-evolve with Their Environment

Core Content: Agent-World is a new open-source framework designed to help AI Agents continuously learn and evolve in real-world environments. The framework supports environmental feedback, action evaluation, and multi-step reasoning, allowing Agents to “grow through experience.”

Commentary: Agent-World addresses the issue of “AI Agents being too dumb”—current Agents are often “one-time use”: you give them a task, they try once, and if they fail, they get stuck. Agent-World enables Agents to learn from environmental feedback, improving with each attempt. This approach mimics human learning: trial, feedback, iteration. It holds significant value for the future applications of Agents.

14. Apple Quietly Cuts Base Mac mini: The Era of “AI Tax” Begins

Core Content: Apple has adjusted its Mac mini product line, eliminating the entry-level configuration, requiring all models to have higher specifications to run Apple Intelligence. This means users must pay extra for AI capabilities, marking Apple’s official entry into the “AI tax” era.

Commentary: Apple’s “AI tax” strategy is clear: AI capabilities are reserved for high-end models, while low-end models are excluded. This is similar to how certain features were reserved for Pro versions of the iPhone. However, AI differs from features like cameras or screens—it is a primary demand, not just an enhancement. Forcing users to pay more for AI may drive them towards Windows PCs, as AI PCs are becoming increasingly capable.

15. Trump Administration Strengthens AI Regulation: White House Plans Pre-release Model Review

Core Content: The Trump administration is considering an executive order requiring AI companies to undergo federal national security reviews before publicly releasing new models. On May 5, Google, Microsoft, and xAI agreed to allow the U.S. government to review models prior to release. This policy shift indicates a transition in U.S. AI regulation from “encouraging development” to a “pre-approval” model, with the White House possibly establishing a new AI task force.

Commentary: This development has far-reaching implications. Previously, AI companies could release products freely; now the government wants to review them in advance. For large companies, this is just an additional step; however, for startups, it could be a matter of life and death—delays in review could derail product timelines. Nonetheless, both the U.S. and China are implementing regulations, albeit with different approaches.

16. Cerebras Launches IPO Roadshow: Target Valuation of $40 Billion

Core Content: AI chip company Cerebras Systems, a competitor to NVIDIA, has officially launched its IPO roadshow, aiming for a valuation of $40 billion. The company expects a 75.7% year-on-year revenue growth by 2025 and to achieve profitability. Cerebras is known for its wafer-scale AI chips, which are 50 times larger than NVIDIA’s, targeting the large model training market.

Commentary: The $40 billion valuation for Cerebras is intriguing—it’s higher than many AI companies, but Cerebras has at least generated revenue and is profitable. Its wafer-scale chip technology is unique and suitable for large model training, but it’s expensive and has high maintenance costs, making it inaccessible for many companies. The success of the IPO will largely depend on whether the market still remembers this company—after all, the concept of wafer-scale chips has been around for several years.