TLDR AI 2025-09-24

Bad actors use AI to test stolen credentials across thousands of sites. They share attack patterns in real-time and automate everything.

Meanwhile, many fraud-prevention teams are still working in silos and are bogged down by manual processes.

The new playbook by Javelin and Plaid looks at how companies can close the gap with modern fraudsters, while making life easier for honest users:

• Connect identity, device, and behavioral signals across the full customer journey
• Share intelligence between institutions to spot repeat offenders
• Monitor patterns across onboarding, login, and transaction activity

Learn how to deliver smarter, more proactive fraud prevention, without sacrificing user experience. Get the playbook.

Apple has begun to lay the groundwork for adopting Anthropic's Model Context Protocol (MCP) in its latest round of betas. MCP allows AI agents to interface with traditional platforms. The standard has been widely adopted and has become a universal pathway for AI assistants to plug into APIs and data sources. Apple plans to let developers use a system-level MCP integration to expose actions and functionalities within apps to AI platforms and agents.

OpenAI has laid out its vision for a $1 trillion build-out of computing warehouses across the US and abroad. It announced five new datacenter sites across the US, built with Oracle and SoftBank, which together will bring online nearly 7 gigawatts of power. The company envisions that it will need more than 20 gigawatts of computing capacity to meet demand for ChatGPT. Each gigawatt is expected to cost roughly $50 billion.

Google Labs' Mixboard is an AI-powered concept board that lets users generate and refine visual ideas using text prompts, image editing, and contextual generation tools.

Perplexity's new Email Assistant for Max subscribers integrates directly with users' inboxes to draft replies, organize messages, and schedule meetings.

AI coding agents are great for new projects or small changes, but they often make developers less productive in large, established codebases. This post discusses a family of techniques called 'frequent intention compaction' that involves deliberately structuring how context is fed to AI throughout the development process to improve performance on large codebases. The technique has allowed a team to use AI to handle Rust codebases with 300,000 lines of code, ship a week's worth of work in a day, and maintain code quality that passes expert review. It shows how AI for coding is actually a deeply technical engineering craft.

Sampling is the process of selecting a token from a model's vocabulary based on the probability distribution. Structured outputs allow models to turn unstructured data into structured data. Together, they both determine a model's next chosen token. Large language models are extremely sensitive to the slightest variation in their system prompts, templating, structured outputs, and sampling parameters.

Running AI on your PC is transforming productivity - but like all emerging tech, it comes with some security risk. Learn about the tactics attackers use to gain entry to your endpoints, and how you can stay secure. Read the eBook by Dell and Intel

GPT-5-Codex is now live in OpenAI's Responses API. It is also available in the Codex CLI via API key. The model has been optimized and is not a drop-in replacement for other models. A link to a prompting guide for the model is available.

The Chrome DevTools Model Context Protocol (MCP) server brings the power of Chrome DevTools to AI coding assistants. It allows AI coding assistants to debug web pages directly in Chrome and benefit from DevTools debugging capabilities and performance insights. Use cases include verifying code changes in real-time, diagnosing network and console errors, simulating user behavior, and automating performance audits. A video showing how the Chrome Dev Tools MCP server works is available.

Google has significantly improved function calling and enhanced proactive audio capabilities in the Gemini Live API to handle interruptions, pauses, and side conversations gracefully. It plans to roll out support for thinking capabilities next week. Details about the recent updates to the Gemini Live API are available in the link along with video demos of the new capabilities.

There has never been a better time in the history of computing to build software. Advances in AI now make it possible to build apps with new capabilities. This post lists a few ideas for builders to consider.

OpenAI wants to create a factory that can produce a gigawatt of new AI infrastructure every week. The project will take several years and require innovation at every level of the stack. A lot of the infrastructure will be built in the US. The company will release more details about its plans and partners over the next couple of months.

Microsoft's Publisher Content Marketplace (PCM) will launch as a pilot program with a limited set of publishers. The company's Copilot assistant will be the first AI buyer on the marketplace. Most AI companies have focused on brokering licensing deals that pay publishers upfront for access to content rather than on a per-use basis. PCM will help Microsoft deepen its relationship with publishers as it expands its AI capabilities and restructures its relationship with OpenAI.

Join this deepset webinar with AI engineer Laura Luckert to see a real-world example of how insurers are using Gen AI to speed up broker support and close more business—without adding headcount. Save Your Seat

The Pew Research Center found that most Americans are increasingly pessimistic about AI's effects on society.

ChatGPT will start using an age-prediction system for under-18 users and stop discussing suicide with teens.

Meta is investing "tens of millions" into the American Technology Excellence Project, a super PAC targeting state-level candidates who oppose AI regulation.