Standard Intelligence has unveiled FDM-1, described as the first fully general computer action model capable of learning real-world digital tasks by training on an unprecedented 11 million hours of video. Unlike typical AI that uses screenshots or short clips, FDM-1 processes long, continuous high-frame-rate video and directly predicts actions (like mouse movements and keystrokes) to interact with software and systems. Early demonstrations reportedly show the model performing complex multi-step workflows such as CAD modeling, automated GUI exploration/fuzzing, and even controlling a car in real-world tests after minimal fine-tuning — suggesting a step toward AI agents that can behave like human users across diverse digital environments.

👉 Read All About It

The U.S. Department of Defense, led by Pete Hegseth, has ordered Anthropic and CEO Dario Amodei to remove key safety guardrails from its Claude AI model for classified military use — or risk losing its $200M contract and potentially facing action under the Defense Production Act. The Pentagon wants unrestricted access for “all lawful purposes,” including capabilities Anthropic says could enable autonomous weapons or domestic surveillance. The company has rejected the demands as unacceptable, setting up a rare and consequential showdown between AI ethics and national security priorities.

👉 Read All About It

A new report claims OpenAI and legendary designer Jony Ive are developing an AI-powered smart speaker with an integrated camera capable of recognizing faces and interpreting its surroundings — marking a shift from command-based assistants to proactive, context-aware systems. Expected to launch in 2027 at a $200–$300 price point, the device could offer features like routine-based suggestions, positioning itself beyond traditional smart speakers and into “observational” AI territory. If accurate, the project signals OpenAI’s ambition to redefine home hardware — but it also raises immediate questions about privacy, comfort, and how far users are willing to let AI see.

👉 Read All About It

Meta has agreed to purchase 6 gigawatts of AI computing power from AMD in a deal worth over $100 billion, potentially giving Meta up to 10% ownership of the chipmaker through milestone-based stock warrants. The agreement centers on AMD’s new MI450 GPUs, custom-optimized for Meta’s AI inference workloads, and marks a major push by AMD to challenge Nvidia’s dominance in AI infrastructure. With Meta planning to spend up to $135 billion this year on AI data centers, the partnership locks in long-term demand for AMD — while highlighting the increasingly aggressive, and sometimes “circular,” financing tactics shaping the AI chip wars.

👉 Read All About It

After a sharp SaaS sell-off and compressed multiples, the “SaaS is dead” narrative is back — but enterprise software spending is actually accelerating into 2026. What’s truly over, Jason Lemkin argues, is the 2010–2024 playbook: win early, outspend competitors with massive sales teams, rely on high NRR and switching costs, and ship incremental updates while customers renew out of inertia. AI-native startups have shattered that model, building agentic-first products in days, scaling to $100M ARR at record speed, and delivering weekly leaps in functionality that legacy vendors can’t match. In this new era, smaller teams, relentless shipping, and undeniable ROI — not lock-in — define the winners.

👉 Read All About It

🤖 TSKR - the intelligent task manager for non-technical founders and chiefs-of-staff (chieves of staff??).

The founder, Eoin, is a friend of the newsletter and a serious operator, coming from Revolut and some other awesome places. Early adoption is, for many, a badge of honour - one that you can don in years to come when the adoptee has found its place in the mainstream such that you can proudly proclaim: “I used that before it was mainstream”. And here you have it folks - an opportunity to adopt early, before TSKR goes to the moon. Great product, great founder, check it out!

Did you know that modern deep learning was quietly rebuilt in London before it went mainstream?

In the early 2000s, when most of the AI world had largely abandoned neural networks, a small group at University College London kept pushing the research forward. At the Gatsby Computational Neuroscience Unit, researchers including Geoffrey Hinton refined probabilistic models and learning techniques that later became foundational to the deep learning systems powering today’s AI.

At the time, this work was considered niche and unfashionable. Just a few years later, the techniques developed and matured in London would underpin the neural networks that now drive everything from ChatGPT-style models to computer vision systems used worldwide.

One of the most important comebacks in tech history was happening in Bloomsbury long before most people realised AI was about to explode. 🤯