Search the archive

you can stay secure. Read the eBook by Dell and Intel

CONTINUING TO BRING YOU OUR LATEST MODELS, WITH AN IMPROVED GEMINI 2.5 FLASH AND FLASH-LITE RELEASE

Google DeepMind has released updated versions of Gemini 2.5 Flash and 2.5 Flash-Lite. The models, available on Google AI Studio and Vertex AI, feature improvements in quality and speed compared to the current stable models. They are also more efficient, with a 50% reduction in output tokens for Gemini 2.5 Flash-Lite and a 24% reduction for Gemini 2.5 Flash. The models are not intended to graduate to a new, stable version, but will help shape future stable releases.

GEMINI ROBOTICS 1.5 BRINGS AI AGENTS INTO THE PHYSICAL WORLD

Google DeepMind has released two models that unlock agentic experiences with advanced thinking. Gemini Robotics 1.5 is a vision-language-action model that turns visual information and instructions into motor commands for a robot to perform a task. The model thinks before taking actions and shows its process and also learns across embodiments. Gemini Robotics-ER 1.5 is a vision-language model that reasons about the physical world, natively calls digital tools, and creates detailed multi-step plans to complete a mission. These advances will help developers build more capable and versatile robots that can actively understand their environments and complete complex multi-step tasks in a general way.

GDPVAL: REAL-WORLD AI BENCHMARKING

OpenAI's GDPval is an evaluation benchmark that tests model performance on real-world, economically valuable tasks across 44 occupations.

BECOMING A RESEARCH ENGINEER AT A BIG LLM LAB - 18 MONTHS OF STRATEGIC JOB HUNTING

Max Mynter signed on as a research engineer with Mistral earlier this week after years of working towards the outcome strategically. This account of Mynter's personal experiences documents how he navigated his career to get where he wanted. During the years, he would work on getting an information advantage and then act on that information to be prepared when it matters. He took the time to learn skills that can't be faked and developed a network of peers and friends who sent opportunities his way.

AI ISN'T REPLACING RADIOLOGISTS

Improvements in image interpretation have run far ahead of their adoption. There are hundreds of models that could potentially help radiologists, but AI is often limited to assistive use on a small subset of scans in any given practice. The promise of AI in radiology is overstated by benchmarks alone. Models can lift productivity, but their implementation depends on behavior, institutions, and incentives. So far, radiologists have only become busier as machines have improved.

Transform your workflow →

SPOTIFY TO LABEL AI MUSIC, FILTER SPAM, AND MORE IN AI POLICY CHANGE

Spotify is adopting a new industry standard called DDEX that identifies how AI was used in a song's production.

COREWEAVE EXPANDS OPENAI PARTNERSHIP WITH $6.5B DEAL, BRINGING TOTAL TO $22.4B

This is CoreWeave's 3rd expansion with OpenAI this year.

IN DEFENSE OF AI EVALS, FOR EVERYONE

Evals aren't about following a rigid philosophy - sometimes it's fine to be less rigorous, and sometimes the task demands more rigor.

OPENAI LAUNCHES SHARED PROJECTS FOR CHATGPT BUSINESS

OpenAI has launched shared projects, smarter connectors, and new compliance controls for ChatGPT Business.

Read the full ebook (online)

OPENAI TESTS CHATGPT AGENT UPGRADES POWERED BY NEW ALPHA MODELS

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.

META'S OPEN LLM FOR CODE AND WORLD MODELING

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'S $7B VALUATION

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

UNLOCKING A MILLION TIMES MORE DATA FOR AI

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.

OPENAI SHOWS US THE MONEY

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.

HUMAN-AI SYNERGY

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.

Dive into the findings >>

CROSS-AGENT PRIVILEGE ESCALATION: WHEN AGENTS FREE EACH OTHER

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.

INTRODUCING THE DATA COMMONS MODEL CONTEXT PROTOCOL (MCP) SERVER: STREAMLINING PUBLIC DATA ACCESS FOR AI DEVELOPERS

CLOUDFLARE VIBE SDK

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.

REINFORCEMENT LEARNING ON PRETRAINING DATA

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

META POACHES OPENAI SCIENTIST TO HELP LEAD AI LAB

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, THE NO. 2 SOCIAL APP ON THE APPLE APP STORE, PAYS USERS TO RECORD THEIR PHONE CALLS AND SELLS DATA TO AI FIRMS

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.

MICROSOFT INVESTS $30 BILLION IN UK TO POWER AI FUTURE

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

MOVIE STUDIO LIONSGATE IS STRUGGLING TO MAKE AI-GENERATED FILMS WITH RUNWAY

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.

INTRODUCING GAMMA 3.0: THE NEW ERA OF HUMAN COMMUNICATION

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

sacrificing user experience. Get the playbook.

APPLE WORKING ON MCP SUPPORT TO ENABLE AGENTIC AI ON MAC, IPHONE, AND IPAD

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 UNVEILS PLANS FOR SEEMINGLY LIMITLESS EXPANSION OF COMPUTING POWER

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'S MIXBOARD FOR VISUAL CONCEPTING

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 EMAIL ASSISTANT LAUNCH

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

GETTING AI TO WORK IN COMPLEX CODEBASES

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 AND STRUCTURED OUTPUTS IN LLMS

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.

you can stay secure. Read the eBook by Dell and Intel

GPT-5-CODEX IS LIVE IN THE RESPONSES API

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.

CHROME DEVTOOLS (MCP) FOR YOUR AI AGENT

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.

BUILD MORE POWERFUL VOICE AGENTS WITH THE GEMINI LIVE API

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.

ABUNDANT INTELLIGENCE

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.

THE 28 AI TOOLS I WISH EXISTED

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.

MICROSOFT LOOKS TO BUILD AI MARKETPLACE FOR PUBLISHERS

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.

Seat