1. Ilya Testifies in OpenAI Trial: Confirms Altman’s Systematic Lying, Holds $7B in Shares
On May 12, 2026, at the federal courthouse in Oakland, California, the Musk v. OpenAI trial reached a pivotal moment as former OpenAI Chief Scientist Ilya Sutskever took the witness stand. Under oath, Ilya confirmed he spent a year collecting evidence of Altman’s “systematic lying,” compiling a 52-page memo submitted to the board. Ilya publicly confirmed for the first time that he holds approximately $7 billion in OpenAI shares (just days earlier, OpenAI President Greg Brockman admitted to holding $30 billion).
Ilya testified that Altman excelled at creating conflicts between executives and manipulating information flow, creating an environment where “executives couldn’t get the right information.” He stated: “I spent a year observing and thinking, and ultimately I believed firing him was the right thing to do.” The trial could determine the future of OpenAI—a company preparing for a trillion-dollar IPO with an $850 billion valuation. Altman is expected to testify as early as Tuesday.
AI Pulse View: Ilya’s testimony is the watershed moment in this “AI trial of the century.” What was previously a commercial dispute between two tech titans has fundamentally changed when OpenAI’s former chief scientist confirmed systemic governance issues as a witness. More concerning is the wealth distribution within OpenAI—just two co-founders hold a combined $37 billion in shares, raising profound questions about wealth distribution during the transition from a nonprofit to a commercial entity. For the AI industry, this case’s outcome could become a landmark precedent for the commercialization governance of nonprofit AI labs.
2. Google Android Show 2026: Called Android, Actually About Gemini’s Ecosystem Rebuild
On May 13, 2026, The Android Show—the prelude to Google I/O—went live online. On the surface an Android launch event, it was actually a full-scenario showcase for Gemini. Google announced four major content areas: Gemini Intelligence, Googlebooks (a new hardware category), Android 17 system updates, and Android Auto updates.
Gemini Intelligence was defined as an umbrella brand—a system that packages all of Google’s AI efforts and clearly delineates “who qualifies to use them.” From Gemini automation to AI-generated desktop widgets, to AI-powered voice input, even Chrome and Android Auto have become Gemini entry points. Google has now split the Android announcement from the I/O main stage for the second consecutive year—meaning Android can no longer stand alongside Gemini models on the main stage, but must have its own stage to carry Gemini’s implementation.
AI Pulse View: Google is taking a completely different AI implementation path from Apple. Apple chose to deeply bind Apple Intelligence to its hardware ecosystem (iPhone/Mac), while Google chose to use the more open but more fragmented Android platform as Gemini’s “hardware backbone.” The Googlebooks category deserves attention—it could be a reading/learning device optimized specifically for Gemini. But the core question remains: Can Gemini’s experience on Android truly surpass the Gemini App on iOS? If Google’s AI strategy ultimately depends on Android for implementation, Android fragmentation could become the biggest obstacle to Gemini’s adoption.
3. Google Confirms: Hackers Already Using AI to Discover and Exploit Zero-Day Vulnerabilities
On May 12, 2026, Google’s Threat Intelligence Group (GTIG) released a report confirming for the first time that criminal hackers used AI large language models to independently discover a previously unknown zero-day vulnerability and wrote a Python script preparing for a large-scale attack. The vulnerability existed in a “widely used open-source web system management tool” and could bypass two-factor authentication (2FA). Google intercepted the attack before any actual damage occurred and notified the relevant vendors.
Google found “AI fingerprints” in the attack code: extensive instructional docstrings (which human hackers would have no reason to include in attack tools), a “hallucinated CVSS score” (a vulnerability severity rating fabricated by AI), and textbook-standard Python formatting. Former NSA cybersecurity chief Rob Joyce called this “the closest thing to a crime scene fingerprint to date.” GTIG’s chief analyst said “this could just be the tip of the iceberg.”
AI Pulse View: The years-long nightmare scenario of “AI automatically finding vulnerabilities” finally has its first confirmed case. The significance of this event is that it proves AI can not only assist human security research but also be used by malicious actors to discover logic-level vulnerabilities that traditional scanning tools struggle to catch. Even more alarming, the hackers likely weren’t using the most advanced models—if current-level AI can independently discover zero-day vulnerabilities, then as model capabilities improve, “AI vs. AI” cyber warfare will enter a whole new speed dimension. Enterprises must rethink their security strategies: from “periodic security audits” to “real-time AI-driven threat detection.”
4. Perceptron Mk1: Video Analysis AI Model Costs 80-90% Less Than GPT-5 and Claude
Two-year-old startup Perceptron Inc. announced the release of its flagship video analysis reasoning model Mk1, priced at $0.15 per million input tokens and $1.50 per million output tokens through its API—80-90% cheaper than Anthropic’s Claude Sonnet 4.5, OpenAI’s GPT-5, and Google’s Gemini 3.1 Pro. Led by Armen Aghajanyan, formerly of Meta FAIR and Microsoft, the model took 16 months to develop.
Mk1 supports native video processing (up to 2 FPS, 32K token context window) and excels across multiple benchmarks: 85.1 on EmbSpatialBench (surpassing Google’s Robotics-ER 1.5 at 78.4), 88.5 on VSI-Bench (the highest among compared models), and 41.4 on the EgoSchema subset (matching Alibaba’s Q3.5-27B). Its “physical reasoning” capability allows the model to understand object dynamics and physical interactions—such as determining whether a basketball shot was taken before or after the buzzer.
AI Pulse View: Perceptron’s strategy is smart—avoiding direct competition with general-purpose models, focusing on the video/physical understanding niche, and opening the market with extreme cost efficiency. An 80-90% cost advantage means enterprises can deploy video AI analysis at scale (sports broadcast auto-clipping, security monitoring, industrial quality inspection) without paying premium prices per call. However, benchmark leadership in video understanding doesn’t equal production reliability—performance in occlusion handling and long-video temporal reasoning scenarios still needs more real-world validation.
5. Andrew Ng Debunks “AI Employment Apocalypse”: Data and Narrative Point in Opposite Directions
On May 13, Andrew Ng published an article arguing that the “AI will cause mass unemployment” narrative has clear stakeholders with vested interests: frontier AI labs need to make their technology sound valuable, AI companies need to shift pricing anchors from software to labor costs, and corporations prefer to attribute layoffs to AI rather than management missteps.
His cited data shows: April 2026 US nonfarm payrolls added 115,000 jobs (far exceeding the expected 55,000), with unemployment stable at 4.3%. A Federal Reserve Bank of Atlanta study found that over 90% of surveyed companies said AI had no substantive impact on their hiring over the past three years. The Yale Budget Lab’s report was more direct: concerns about AI employment impact “currently remain largely speculative.” Ng used VisiCalc as a historical example—the spreadsheet software was predicted to cause mass accountant unemployment, yet the number of accountants grew 4x over the next 40 years (Jevons Paradox).
AI Pulse View: Ng’s analysis cuts to the three core interest chains of “AI employment panic,” which is an important correction. However, it’s worth noting that his arguments rely primarily on macroeconomic data, while AI’s impact on employment is often structural—some jobs are replaced, others are created, and macro data may mask this structural change. The real question isn’t “will total employment decrease?” but “is there a skills gap between replaced and created jobs?” For policymakers, the focus should be on skills training and transition support, not simply denying or exaggerating AI’s employment impact.
Other Developments
- DeepSeek raises $7B: According to 36Kr, DeepSeek is conducting a massive funding round. Liang Wenfeng faces capital maneuvering, with core demands for both capital and control.
- Elsevier sues Meta: Elsevier accuses Meta of using pirated papers from Sci-Hub to train large language models; Meta defends with “fair use.”
- Unitree CEO visits China Huaneng: Energy SOEs intensify contact with embodied AI companies, promoting bidirectional computing-power synergy. Unitree’s manned transforming mecha priced at 3.9 million yuan.
- OpenAI former CTO’s startup debuts: New company’s first model directly challenges ChatGPT’s chatbox experience, emphasizing “presence” as a core interaction philosophy.