TLDR AI 2025-05-13

AI agents and MCP servers are powerful—but without proper access control, they can become liabilities. When agents and services have unchecked access, data exposure and unintended actions are inevitable.

Enter WorkOS AuthKit—your authorization layer for AI agents and services. With scoped OAuth access and built-in role-based access control, you can assign roles, define permissions, and lock down exactly what each service can access.

Lock down access today with WorkOS AuthKit.

OpenAI's $500 million Stargate data center initiative is reportedly struggling due to investor hesitancy tied to economic volatility.

Google's AI Futures Fund will invest in startups using DeepMind's AI tools and offer access to models, cloud credits, expert support, and potential direct funding.

Google is reportedly set to unveil an AI software development agent that assists with the complete development lifecycle at its I/O conference on May 20. The move puts Google in direct competition with Anthropic's Claude Code, OpenAI's Windsurf, and numerous startups in the increasingly crowded AI coding space. There may be potential Gemini integration with AR glasses.

Amazon's custom stowing robot matches human performance in warehouse operations while highlighting robotics' current frontier - its specialized hardware and AI vision can successfully handle diverse items at scale, yet the 14% failure rate demonstrates why full warehouse automation remains elusive despite significant advances.

Hugging Face has outlined how Vision Language Models have advanced with smaller, more capable architectures, enabling reasoning, video understanding, and multimodal agents.

China's AI infrastructure boom has led to severe overcapacity, with 80% of computing resources across 500+ new data centers sitting unused. DeepSeek's R1 model fundamentally changed market dynamics by shifting demand from training-focused to inference-optimized hardware, rendering many facilities obsolete. Despite this market correction, China has continued to aggressively invest in infrastructure to compete with U.S. initiatives like the $500 billion Stargate project.

UniVLA enables policy learning from unlabeled video across diverse robot embodiments by inferring task-centric latent actions.

OpenAI's HealthBench is a benchmark developed with 262 physicians for evaluating AI models on realistic medical conversations.

LLMs perform significantly worse in multi-turn conversations, with an average 39% drop in task performance due to unreliability and early, incorrect assumptions.

IBM, Princeton, CMU, and UIUC released Bamba v2, a Mamba2-based model that outperforms Llama 3.1 8B after training on 3 trillion tokens. Bamba v2 leverages Mamba2 architecture for 2-2.5x faster inference and excels in L1 and L2 benchmarks. The team plans further optimizations and invites community collaboration to enhance model performance.

Figma has introduced Figma Sites, allowing users to design, build, and publish responsive websites directly from Figma, streamlining the design-to-production workflow.

A survey of media professionals from Reuters, The Washington Post, VentureBeat, 404 Media, and others shows newsrooms are adopting AI selectively, embracing it for transcription, data analysis, and translation while rejecting AI content creation. While Reuters reports that AI now generates about 25% of its code, many journalists remain skeptical, prioritizing maintaining audience trust and journalistic integrity over technological efficiency.

Manus has eliminated its waitlist, offering broader access to its virtual desktop AI agent with one free daily task for all users and a one-time bonus of 1,000 credits and significantly lowering the barrier to entry for this previously hyped automation tool.

Helium 1, a 2 billion parameter LLM, excels in European languages and is optimized for on-device use.

The Trump administration has canceled Biden's "AI diffusion rule," which would have restricted American technology exports.