AI that matters, from the architect's desk. Curated and engineered by Saaket Varma, PhD — no hype, just signal.
HOW TO READ THIS Read top to bottom: Meta moves from AI maker to compute seller, offering compute and models for rent, entering the row of clouds beside AWS, Google, and Azure, which now form a four-way market.
Meta is drawing up plans to sell access to its AI compute and models, putting it in direct competition with AWS, Google Cloud, and Azure. This is the clearest sign yet that the AI capex arms race is becoming a revenue line: infrastructure built for a company's own models turns into a product the moment there's capacity to spare. For buyers, a fourth serious seller of AI compute changes negotiating leverage — every cloud pricing conversation now has one more credible alternative in the room. If you're locking in multi-year GPU or inference commitments, price this into the negotiation before you sign.
HOW TO READ THIS Read top to bottom: two labs act, each ships a tool, the tools converge on the same job, and the payoff is still research-stage.
Anthropic launched Claude Science, an AI workbench built for scientists, the same week OpenAI introduced GeneBench-Pro, a benchmark testing models against real-world genomics and biology datasets. Two frontier labs converging on the same surface in one week means science tooling has moved from demo to product line. If your work touches research, data analysis, or lab workflows, this is the moment to evaluate purpose-built tooling instead of forcing a general chat model into the job — and benchmarks built on real datasets finally give you something concrete to hold vendors to.
HOW TO READ THIS Read top to bottom: Alibaba builds it, ships it inside a webpage, the agent inside acts directly on page elements with no screenshot step, and the repo hits 20.9k stars.
Alibaba's page-agent — top trending on GitHub at roughly 21k stars — is a TypeScript GUI agent that runs inside the webpage itself and drives the interface from natural language by operating the DOM directly, not by screenshotting and clicking. That's a materially lighter pattern than conventional browser automation: no separate browser process, no vision model in the loop. If you ship a web app, this is the cheapest path yet to adding a 'just tell it what to do' layer — worth a prototype before committing to a heavier computer-use stack.
HOW TO READ THIS Read top to bottom: OpenAI teases a device on social media, that device is revealed to house the Codex coding agent, and the full reveal lands July 15.
OpenAI posted a video of a square, button-covered device tied to its Codex coding tool, with a reveal set for July 15. A dedicated physical gadget for a coding agent is a first, and it signals where OpenAI thinks agentic coding is headed: out of the terminal window and into its own interface. Nothing to act on until the reveal — but if coding agents are getting purpose-built hardware, the 'agent as a background coworker' idea is being taken literally.
An agent multiplexer that lives in your terminal — run and switch between multiple coding agents from one place.
Battle-tested Node.js web scraping and browser automation library for building reliable crawlers — the data-collection layer under many agent stacks.
macOS dictation with fully on-device speech-to-text plus a custom AI enhancement model — a local alternative to Wispr Flow.
All-in-one agentic framework for video understanding, editing, and remaking.
Self-hosted, user-friendly AI interface that fronts Ollama and OpenAI-compatible APIs — the default pick for local model UIs.
Latent Space How Cursor deploys AI inside the enterprise — forward-deployed engineers as the delivery model for agentic tooling.
Ben's Bites GPT-5.6 lands — a first read on what shipped, with the caveats the headline's 'but' promises.