Back to list
Daily 2026-05-14

AI Pulse Daily | 2026-05-14

Tian Yudong StartupBaidu Create2026AnthropicAI CopyrightGemini

1. Former Meta AI Scientist Tian Yudong Co-founds Recursive Superintelligence with $650M Funding

Tian Yudong, former Meta AI scientist, officially announced his role as co-founder of Recursive Superintelligence. The secretive startup went public today alongside a $650 million funding round at a $4.65 billion valuation. GV (Google Ventures) and Greycroft led the round, with AMD Ventures and NVIDIA participating. The company has 8 co-founders from OpenAI, Google DeepMind, Meta AI, Salesforce AI, and others, betting on “Recursive Self-Improvement” — enabling AI to autonomously discover knowledge and continuously optimize itself. AI pioneer Peter Norvig has also joined the team.

AI Pulse View: Tian’s decision to reject offers from OpenAI, XAI, and Anthropic in favor of starting up signals strong conviction in the Recursive Self-Improvement direction. With $650M in funding and an all-star founding team, this could be one of the most noteworthy AI startups of 2026.

Source: 36Kr | 2026-05-14

2. Baidu Create2026: Robin Li Introduces “Daily Active Agents” (DAA) Metric

At the Create2026 Baidu AI Developer Conference in Beijing, CEO Robin Li introduced “Daily Active Agents” (DAA) as the new metric for the AI agent era. He argued that tokens represent cost, not value — DAA measures how many agents are actually delivering results for humans. Li predicts global DAA could exceed 10 billion. Baidu also launched DuMate mobile app, Miaoda (code agent) app & enterprise edition, upgraded Huiboxing to “Baidu Yijing” digital human platform, and released Famo 2.0 decision-making agent.

AI Pulse View: Migrating the DAU concept to the agent world is an interesting metric innovation. DAA could become a key indicator for evaluating AI product commercial value. Baidu’s full-stack “chip-cloud-model-agent” layout shows it’s building a complete ecosystem from compute to agents.

Source: QbitAI | 2026-05-13

3. Anthropic Surpasses OpenAI in Enterprise AI Adoption for the First Time

VentureBeat reports that for the first time, more American businesses are paying for Anthropic’s Claude than for OpenAI’s ChatGPT, marking a significant shift in the enterprise AI landscape. However, analysis points to three major threats that could erode Anthropic’s lead.

AI Pulse View: Claude’s rise in the enterprise market reflects developers’ emphasis on API friendliness and tool ecosystems. However, OpenAI’s massive consumer-side advantage remains intact. The enterprise market’s tipping point will depend on who provides more reliable Agent infrastructure.

Source: VentureBeat | 2026-05-13

4. Anthropic Reinstates OpenClaw and Third-Party Agent Support with New Credit System

Anthropic has re-allowed OpenClaw and third-party agent usage on Claude subscriptions, but introduced a new Agent SDK credit system ($20-$200). Inefficient agents will consume the user’s credit budget faster rather than exceeding the value of Anthropic’s fixed subscription tiers.

AI Pulse View: This strategy cleverly shifts the agent efficiency problem to users rather than the platform — users will naturally phase out inefficient agents, while Anthropic retains predictable subscription revenue. This is an important experiment in the agent economy model.

Source: VentureBeat | 2026-05-13

5. Frontier AI Models Don’t Just Delete Content — They Rewrite It Silently

VentureBeat reports that while weaker AI models delete document content when they fail, frontier models tend to silently rewrite it — making errors far more subtle and nearly impossible for human reviewers to catch.

AI Pulse View: This is a concerning security issue. Silent rewriting is far more deceptive than deletion, and in high-stakes scenarios like finance and law, this could trigger serious compliance risks. Future AI outputs will need stronger traceability and verification mechanisms.

Source: VentureBeat | 2026-05-13

6. Elsevier Sues Meta: Using Sci-Hub Pirated Papers to Train Llama Models

Academic publishing giant Elsevier, together with Hachette (France), Macmillan (UK), and novelist Scott Turow, has filed a lawsuit against Meta and CEO Mark Zuckerberg in the Southern District of New York. The complaint alleges Meta illegally obtained copyrighted academic papers via Common Crawl and piracy platforms Sci-Hub/LibGen to train its Llama models. Meta says it will vigorously defend, citing “fair use.”

AI Pulse View: This is the first time top academic publishers have sued an AI company over copyright, marking a new phase in the AI copyright debate. While US courts previously recognized “fair use” in the Anthropic case, academic papers may enjoy stronger copyright protection. This case will have far-reaching implications for the entire industry.

Source: QbitAI | 2026-05-13

7. MiniMax Launches Mavis: Agent Management System with Hierarchical Governance

MiniMax has launched Mavis, a new agent management system that uses a hierarchical governance structure (likened to ancient Chinese bureaucratic systems) to manage multiple agents. The system aims to enable “managing agents like managing people” with support for scheduling, collaboration, and permission management across multi-agent teams.

AI Pulse View: As agent counts grow, management architecture becomes essential. Mavis’s hierarchical approach borrows from traditional organizational management metaphors, but the core challenge remains making inter-agent communication and task allocation more efficient.

Source: 36Kr | 2026-05-13

8. Google Gemini Goes All-In Across Its Ecosystem

Google has fully integrated Gemini AI capabilities across the Android ecosystem, including system-level AI assistants, input suggestions, and image recognition. Google is the first to realize the AI ecosystem vision that Apple outlined at WWDC.

AI Pulse View: Google’s execution on system-level AI integration is ahead, leveraging Android’s openness and Google’s AI infrastructure. However, Apple’s unique advantages in privacy protection and on-device inference remain strong — WWDC 2026 will be worth watching.

Source: QbitAI | 2026-05-13

9. ByteDance Proposes Third Route for Visual Generation

ByteDance researchers have proposed a new approach to visual generation, challenging the two dominant paradigms: diffusion models and autoregressive models. The new method lets models “draw and revise simultaneously,” like a human artist, enabling dynamic adjustment and optimization during the generation process.

AI Pulse View: Visual generation has long been dominated by diffusion models, with autoregressive approaches gaining ground. If ByteDance’s new route can achieve breakthroughs in quality and efficiency, it would provide a third option for AI image/video generation, especially for scenarios requiring precise control.

Source: QbitAI | 2026-05-13

Other Updates