1/ Today we're launching Cua Driver Bench, together with Scale AI's AgentEnv. Can an agent do a desktop task in the background while you keep working undisturbed? On macOS, Cua Driver fully passed 58%, Codex computer use 33%, Vercel's agent-desktop 20%, browser-use 3%.

I announced earlier this week that I am leaving Meta. I am deeply grateful to Mark Zuckerberg for showing what a truly founder-led company can achieve. I am incredibly optimistic for Meta’s future. At Meta, I was able to witness the most exciting opportunities in enterprise AI. That’s why today, I’m announcing that I will be joining Fireworks as CFO. I am thrilled to be joining a founder-led growth company where I can roll up my sleeves and put my experience to work. Fireworks has the right mission for the moment. Companies will use AI to build on what makes them distinctive: their data and expertise. Specialized intelligence gives them a way to capture those advantages and own their “alpha.” Open models offer price points that allow the technology to have the greatest impact. Fireworks is building the platform to make all of this possible. I’ve known and respected Lin @lqiao for a long time. As I’ve spent more time with her and the people at Fireworks, I’ve become increasingly bullish on what they are building. There’s a remarkable depth of technical talent here, coupled with a very practical focus on what customers need. I see AI as the most consequential technology shift of my lifetime. I’m excited to get to work and help build a business at the center of it.

Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time.

hahaha Claude Code dopisywał Twój adres e-mail do nagłówka User-Agent, czyli przekazywał go czasami stronom www, które twój agent analizował/kradł albo odpytywał o informacje potrzebne do twojego projektu. Efekt? Użytkownicy Claude Code zauważyli że oststnio zaczęli dostawać "podejrzanie sprofilowany spam". Anthropic potwierdził, że wyciek adresów e-mail przypisanych do kont to wynik uboczny jednej z poprawek. I że naprawią to w kolejnej poprawce 🤡

Claude Duet: now with Claude Code sessions Check what's running, approve changes and send updates - all from your phone. Works with sessions on your computer or in the cloud. /cc @navbryce and @joshleibsly

Playing around with spaces in Muse lately -- tell Muse what you need and it builds it. I've made a BBQ game, an event tracker, and a home control dashboard. Excited to see what you build.

1/ starting today, muse and meta ai will run on new specialized models trained using muse wonder. ACTIVE MEMORY: models can reason across time and better remember what matters ACTIVE FOCUS: enabling muse to continue working towards a goal even when met with roadblocks

We've made it easier to get your repo running in Claude Code cloud sessions on Pro, Max, and Ultra. Claude figures out what dependencies, network access, and API keys it needs, then saves the environment for next time. API keys are saved securely, so Claude never sees them.

Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line. After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs. The values in an open model are one fine-tune away. Our eval, weights, and training data are available now along with the full blog post:

I spent 20,000 dollars in API credits having different AI models play each other at Blood on the Clocktower (a Mafia or Werewolf type social deception game). The results were fascinating. Models have gotten much better over the course of this year at creative on-the-spot scheming and bluffing, and also at reasoning as a team to break apart these underground conspiracies. I go through some of the most standout games these AIs played in this video. The new models are getting pretty good. If my friends pulled off some of these moves while we were playing, I’d give them a fist bump. Fwiw, OpenAI’s newer models are much better than Anthropic’s at this game. You can see the full benchmark at the website, and I’ve also made it so that you and your friends can play again the AIs:

Two months ago, we introduced Atlas, an autoregressive world model with pixel-perfect camera control. Today, we're releasing Frameboy, a demo allowing you to explore worlds in real-time with a highly optimized version of Atlas, streaming from cloud GPUs at up to 15 fps.

You can now build live apps in Cursor and share them with a link. Agents have made it easier to write code, but most apps never make it off localhost. Describe the tool you want - Cursor will build it, set up a database, run deterministic and agentic security checks, and deploy.

New research: We developed JOLT, a post-training method where one model acts as both teacher and student, with the teacher trained to give feedback the student can use. On MATH, it matched GRPO's accuracy with 6.5x fewer training tokens.

New: I spoke with @sundarpichai about Google's commitment to the AI frontier, how close we are to RSI, balancing DeepMind's compute needs vs. Google Cloud profits, and much, much more. Full Big Technology Podcast episode here --> Chapters: 0:00 Introduction 1:14 Google’s Frontier AI Strategy 3:47 Do Frontier Models Really Matter? 6:35 The Future of AI Agents 7:39 Recursive Self-Improvement 12:58 World Models and the Path to AGI 14:56 Google’s Compute Dilemma and AI Infrastructure 19:36 Is AI Overbuilt? 23:35 The Future of the AI Interface 31:42 Google’s Answer to Muse and Dots 35:03 Apple, Space Data Centers, and AI Coding 41:42 Optimism and AI’s Future