Anthropic launched Claude's computer use for Mac — enabling Claude to physically point, click, scroll, and navigate across any app on a user's desktop.
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.
Anthropic introduced its Science Blog, positioning scientific acceleration as a core part of the company's mission alongside safety.
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.
Ethan Mollick observed a sharp philosophical split between OpenAI's Codex and Claude Code skills — Codex skills are functional references, while Claude skills teach the AI approaches to problems.
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.
Wes Roth reported that Anthropic is quietly building a dedicated Projects layer for Claude Cowork, with local folders and project-specific scheduling.
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.
Sahil Lavingia turned The Minimalist Entrepreneur into 9 Claude Code skills — from /find-community to /marketing-plan — encoding an entire business philosophy into executable AI workflows.
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.
Jeff Clune's team released HyperAgents — a self-improving AI system that extends the Darwin Gödel Machine by improving not just task performance but the very mechanism by which it improves.
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.
Mckay Wrigley is privately testing a new tool aimed at Claude Cowork, OpenClaw, and Manus users — with a 4-hour self-destruct timer on his DM call.
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.
A photo of a Mac mini farm — racks of Apple silicon boxes — surfaced with the observation that you can run OpenClaw AI agents on these all day, "or it may be just bot farm."
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.
From Rakesh B · 52K views, 339 likes
Jensen Huang told Lex Fridman "I think we've achieved AGI" — the most direct claim of its kind from a major industry leader, on the most prominent AI interview platform.
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.
Jensen Huang separately urged that every college student should become an AI expert, framing AI literacy as a universal requirement for the modern economy.
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.
Perplexity Computer can now access real-time options chains and execute live trades via a free API — pulling quotes, scanning chains, checking delta/IV, and placing orders through conversation.
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.
OpenArt shipped a browser-based 3D world builder that generates navigable environments from a single prompt or image — one prompt, and you're walking through a rendered world in seconds.
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.
CapCut launched Dreamina Seedance 2.0 — a generate-and-edit video tool rolling out across Southeast Asia, Brazil, and Mexico first.
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.
MirrorMirror AI launched a marketplace where fashion models license their likeness for AI-generated imagery — creating a new revenue stream from identity as digital asset.
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.
Iran issued a sweeping ultimatum to Washington — demanding closure of all US bases, a $2M toll for Western ships in Hormuz, and $100 billion in reparations — drawing 1.1 million views.
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.
Tuki compiled a single-Monday summary: Zuckerberg building AI to do his CEO job after firing 31K people, while someone bought $1.5B in S&P futures five minutes before Trump halted Iran strikes.
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.
Garry Tan declared that in the AI revolution, "low status and useful is where the alpha is" — championing markdown files as the ultimate human-AI interface.
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.
Vitrupo cited Wolfram's observation that AI agents are least useful where expertise is highest — for experts, thought-to-code is already faster than thought-to-English.
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.
Shankar Padmanabhan presented DiSC, a method for updating LLMs with new world knowledge without undoing skills learned during post-training.
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.
Today's feed is dominated by a single narrative — AI leaving the browser and arriving on the desktop. Claude's computer use announcement, Cowork trending, Jensen's AGI declaration, Qwen on iPhone, Sahil encoding a whole book into executable skills. The common thread isn't any one product launch. It's the moment where AI stops being something you visit and becomes something that lives with you.