TLDR AI 2025-09-25

Chatbots might be the current focus of attention — but under the radar, genAI is completely revolutionizing search.

Algolia's new report reveals how GenAI transforms the entire search stack from content lakes and vector databases to RAG implementations and content generation systems.

The guide explores how NeuralSearch engines now combine multiple search types, why retrieval-augmented generation (RAG) is becoming essential for quality control, and how companies are rebuilding their search infrastructure around LLM orchestration.

Also covered: breakdowns of evaluation frameworks, chunking strategies, and the headless architecture patterns that leading teams use to integrate genAI without breaking existing systems.

Read the full ebook (online)

Meta has released CWM, a 32B decoder-only LLM trained on code execution traces and reasoning tasks to explore world models in code generation.

Cohere secured another $100 million in funding, raising its valuation to $7 billion, and announced a strategic partnership with AMD.

Some ChatGPT users have spotted a new 'alpha models' section in the model selector. The models appeared for a limited time, and they activated agent mode. The model naming - 'Agent with truncation' and 'Agent with prompt expansion' - suggests that OpenAI may be experimenting with different system prompt setups or underlying model architectures. The release seemed accidental and was quickly rolled back, making model comparisons challenging. The new agent variants appear to be aimed at early adopters, researchers, and power users interested in tool-augmented AI workflows.

Nvidia will invest as much as $100 billion in cash into OpenAI to support new data centers. As long as OpenAI depends on Nvidia, it may be reluctant to speak about AI risk, as it may endanger its chip allocations. While Nvidia wants to make and sell as many chips as possible to the widest variety of customers, OpenAI has a strong interest in the US beating China across the board and keeping things under control. Nvidia may do better short-term by selling to China, but this will hurt other investments, and long-term, the model of compute-intensive, top-quality closed models is better for it anyway.

Every major leap forward in AI progress involved a large increase in data to support it. While AI leaders have warned that we have reached 'peak data', there is still a lot of unused data available. Most AI systems are trained on a few hundred terabytes of data - the world has digitized an estimated up to 200 zettabytes of data. The data exists, but it isn't being used for training, which is an access problem rather than a scarcity problem. The Attribution-Based Control is a potential framework for expanding access to the world's digital data while preserving ownership rights. It offers an approach to addressing the technical, legal, and economic barriers that prevent access to most of the world's data by enabling data owners to maintain control while contributing to AI development.

Across 600 participants, researchers found that people who better understand others' mental states performed significantly better with AI assistance. The findings suggest that successful human-AI collaboration depends less on technical prowess and more on social intelligence.

Despite billions invested, most AI pilots never scale and the reason is clear. Traditional monitoring tools were not designed for the complexity, unpredictability, and autonomy of today's AI systems.

In this research, discover how observability closes this critical AI production gap. From monitoring hallucinations and tracing decision paths to ensuring compliance and cost optimization, observability gives enterprises the visibility and control needed to deliver trustworthy AI outcomes at scale.

Dive into the findings >>

Cross-Agent Privilege Escalation is when multiple coding agents operating on the same system are tricked into modifying each other's configurations to escalate their privileges. While some agents now lock down the ability to modify their own settings, this doesn't work when multiple agents are run in the same environment. The possibility of the exploit highlights the need for better isolation strategies and stronger secure defaults in agent tooling. Developers should get into the habit of running these agents in locked-down containers.

The Data Commons Model Context Protocol Server makes all the Data Commons' vast and interconnected public datasets instantly accessible and actionable for AI developers, data scientists, and organizations worldwide. It allows developers to use real-world statistical information to help reduce large language model hallucinations. The server enables agents to handle the full range of data-driven queries, from initial discovery to generative reports. It is designed for seamless integration into agent development workflows.

The Cloudflare Vibe SDK is an open-source full-stack AI webapp generator. It deploys an AI-powered platform that allows users to specify what they want to build in natural language, and the AI creates and deploys that application. The SDK is perfect for companies that build AI-powered platforms, internal development, and SaaS platforms. Built on Cloudflare's platform, the Cloudflare Vibe SDK utilizes the full Cloudflare developer ecosystem. A full guide on how to set it up is available.

RLPT introduces a new training paradigm where language models use reinforcement learning directly on pretraining data, removing the need for human-annotated rewards.

Mark Zuckerberg has poached Yang Song, a high-ranking researcher from OpenAI, to be the lead research principal of Meta Superintelligence Labs. Song had been at OpenAI since 2022, where his research focused on improving models' ability to process large, complex datasets across different modalities. He developed a breakthrough technique that helped inform the development of the DALL-E 2 image generation model while still a graduate student at Stanford University.

Neon Mobile pays users for their audio conversations, selling the data to AI companies. It pays 30 cents per minute when users call other Neon users and up to $30 per day maximum for making calls to anyone else. Only Neon users are recorded during these calls. The app also pays for referrals. The app's privacy policy includes a very broad license to its user data that leaves plenty more room for the company to do with users' data than it claims. While the app raises many red flags, it may be technically legal.

Clockwise MCP helps AI schedule time like a human would. Schedule meetings, optimize your work calendar, and intelligently block time for tasks. Available today in Claude, Cursor, Relay, and more.

Lionsgate doesn't have enough content for Runway's AI to produce quality content - even Disney's catalog is too small to create a model.

Gamma is an AI-powered design platform that transforms raw ideas into compelling visual presentations.

Microsoft is investing $30 billion to enhance AI infrastructure in the UK from 2025 to 2028, including a major supercomputer project with Nscale.