Monday Evening Edition

The feature lets Claude operate a user's machine the way a human sitting at a desk would: opening apps, filling spreadsheets, navigating browsers. This is Anthropic's clearest move yet toward making Claude a persistent desktop co-worker rather than a chatbot in a tab. "Claude Cowork" was trending in Technology at the time of this scan.

The blog will feature new research and stories of how scientists use AI to accelerate their work. Anthropic is diversifying its public identity beyond safety research toward being seen as a platform for scientific discovery, mirroring Dario Amassei's earlier essay on the intelligence explosion's benefits.

The distinction is revealing: OpenAI's skills tell the model what to do; Anthropic's skills tell it how to think. This maps neatly onto each company's broader identity — one engineering-first, the other more interpretive and adaptive.

Independently confirmed by TestingCatalog News. The move extends Claude Cowork from a per-session tool into a persistent project management surface — the kind of feature that turns occasional use into daily habit.

This is a proof-of-concept for a new form factor: books as code. Rather than reading a chapter on pricing, you invoke /pricing and get an interactive analysis adapted to your specific product. The Gumroad founder is effectively demonstrating that the next generation of business books may ship as skills, not chapters.

Co-authored across UBC, Vector Institute, Edinburgh, NYU, and Meta, HyperAgents introduces self-referential agents integrating a task agent and meta-agent into a single editable program. By making the improvement procedure itself editable, the system breaks the assumption that task performance and self-modification skill must be aligned — enabling self-acceleration across coding, paper review, robotics, and Olympiad math.

In December he declared you're "6 prompts away from infinitely customizable personal AGI." The scarcity play is deliberate, but the signal is real: the personal-agent tooling layer is getting crowded fast, and builders are racing to own the orchestration surface above the model.

The ambiguity is the point: the hardware cost of running persistent AI agents has collapsed to where a shelf of Mac minis becomes a plausible agentic computing cluster. This is infrastructure as still life — the visual equivalent of the server rack, but in someone's apartment.

Lex's framing was characteristically maximalist: the definition he offered was an AI system capable of doing Jensen's own job. Jensen's agreement signals not just confidence in current capabilities but a willingness to shift the Overton window on what counts as general intelligence, likely to accelerate the AGI narrative across the industry.

The statement positions Nvidia as an implicit education platform — its hardware enables the models that every student will need to learn to use. Coming on the same day as his AGI claim, it reads as a one-two punch: AGI is here, and you need to learn how to work with it.

This is a Mixture-of-Experts model with 256 experts running fully on-device at conversational speed. The upstream enabler is Apple's increasingly capable Neural Engine combined with aggressive kernel-level optimization. A separate post showed the same model at 5.6 tok/sec on a previous iPhone, making the 13.1 figure a genuine leap in on-device capability.

This is the first major AI product to integrate real-time financial instrument execution natively, turning a search engine into a brokerage interface. The regulatory and risk implications are significant: conversational options trading eliminates the friction that previously served as a natural guardrail.

The XR and spatial computing implications are immediate — this collapses the 3D content creation pipeline from days of modeling to seconds of generation. The browser-native approach means no download, no install, just creation.

The regional rollout strategy is notable: rather than US-first, ByteDance is seeding its most powerful creative tools in markets where TikTok's creator ecosystem is growing fastest. This is distribution-first product strategy at continental scale.

The approach sidesteps the consent crisis in AI image generation by building licensing into the foundation rather than bolting it on after the fact. Commercially licensed, model-approved — a business model that might actually survive the coming regulatory wave.

Trump subsequently postponed planned strikes on Iranian power plants after "very good" talks, extending the Strait of Hormuz deadline. The sequence suggests a high-wire diplomatic dance where maximalist demands create negotiating room.

The juxtaposition captures the surreal velocity of concurrent crises — each item would dominate a normal news cycle, but today they're competing for the same Monday. The $1.5B futures trade timing, if accurate, is the kind of data point that spawns investigations.

His argument: markdown has zero glamour but is readable by humans and models alike, writable, diffable, transformable, and chainable. The observation names something power users of Claude Code already know intuitively — the most valuable formats in an AI-mediated workflow are the ones nobody would put in a pitch deck.

The implication: AI's greatest impact may be in adjacent competency zones — not replacing experts in their core domain, but making them dangerous in domains they'd otherwise never enter. AI doesn't replace the best; it makes everyone else competitive.

The paper addresses a real pain point: continued training on new facts typically degrades instruction-following and math abilities. DiSC uses document splitting to inject knowledge while preserving capabilities — a quiet but important infrastructure advance for keeping models current.