TLDR AI 2025-05-22

Your app isn't just meant for humans anymore. AI agents, APIs, and partner applications are waiting for access… but are your APIs ready?

Use Descope to make your APIs and remote MCP servers OAuth-compliant with a few lines of code. Turn your app into an identity provider and enable seamless AI agent connectivity.

🤖 AI agents get secure and scoped access to resources.

🙋‍♀️ Users get visibility and consent mechanisms.

🫵 You get to securely open up your product to the agentic world.

Databricks, You .com, GoFundMe, and leading AI companies trust Descope. Make your app AI-ready today

OpenAI has fully acquired io, its joint venture with famed Apple designer Jony Ive, for $6.5 billion in equity as it moves into hardware development.

Mistral AI and All Hands AI have introduced Devstral, a new open-source LLM optimized for software engineering.

A toy example demonstrating the basics of Minimum Cost Path (MCP) and Agent-to-Agent (A2A) ping checks.

Tech hiring for new graduates has collapsed by over 50% since pre-pandemic levels while experienced engineers remain in high demand as AI tools increasingly handle entry-level tasks. Anthropic has emerged as the clear leader in the AI talent wars, maintaining an impressive 80% employee retention rate while aggressively poaching engineers from competitors. Talent is now 8 times more likely to leave OpenAI or DeepMind for Anthropic than move in the reverse direction.

Large language models exhibit strong capabilities in natural language reasoning about visual imperfections, but they struggle to map these high-level insights into precise pixel-space action. The models excel when reasoning is constricted to discreet, well-scoped dimensions - performance degrades when they are asked to balance abstract aesthetic judgments with deterministic pixel-level corrections. This suggests that the models have trouble bridging symbolic understanding and spatial grounding, especially in iterative workflows that demand surgical image edits.

The world's biggest event for AI builders is in SF from June 3-5. Speakers include leaders from OpenAI, DeepMind, Anthropic, and more, click here to get 30% off your ticket.

Google DeepMind's research on defending Gemini against indirect prompt injection attacks reveals that more capable models aren't inherently more secure, and defenses that work in static tests often fail against adaptive attackers. Their findings show that "adversarial training" - fine-tuning models on examples of malicious prompts - significantly improved Gemini 2.5's resistance without degrading performance on normal tasks. When they combined this approach with a "Warning" defense, attack success rates plummeted from 94.6% to just 6.2%.

Giving large language models the full output of tool calls is costly and slow. Output schemas enable developers to get structured data, which can then be processed. Using code execution as a way to process data from MCP tools opens the door to scaling the work of AI models. However, allowing an execution environment to also access MCPs, tools, and user data requires careful design as to where API keys are stored and how tools are exposed.

OpenAI has added built-in tools and features to the Responses API. The API now has support for all remote Model Context Protocol (MCP) servers and tools like image generation, Code Interpreter, and improvements to file search. It has new features that improve reliability, visibility, and privacy for enterprises and developers. All of these tools and features are now available in the Responses API, which supports the GPT‑4o series, GPT‑4.1 series, and o-series reasoning models. Image generation is only supported on o3 in the reasoning model series.

AI products depend heavily on user prompts, unlike traditional software with predictable outcomes. Skilled prompting can yield varying results due to nuanced intent and context. Product teams can improve results by refining prompts and using follow-up questions to guide users.

FutureHouse combined specialized literature search agents with a data analysis agent in a continuous experimental loop to accelerate medical discovery. The system autonomously generates hypotheses from literature reviews, proposes experiments for human execution, then analyzes resulting data to inform the next research cycle. Through this process, researchers identified ripasudil, an eye drop medication that relaxes cellular tension, as a promising new treatment for the leading cause of age-related vision loss, where retinal light-sensing cells gradually deteriorate. The complete code, data, and agent interaction traces will be publicly released on May 27.

v0's AI model, which has specialized web-dev knowledge and an OpenAI-compatible API, is now in beta in the API, AI SDK, and AI Playground.

Extended chain-of-thought reasoning helps large language models better calibrate their confidence.

LM Arena, a crowdsourced AI benchmarking platform, has secured $100 million in seed funding led by Andreessen Horowitz and UC Investments, despite recent accusations from some researchers that it helps top AI labs game its influential leaderboards.

Meta has announced a new initiative to support early-stage U.S. startups using its Llama models.

Gemini Diffusion is Google's first large language model to use diffusion in place of transformers - it has the performance of Gemini 2.0 Flash-Lite at five times the speed.