AI that matters, from the architect's desk. Curated and engineered by Saaket Varma, PhD — no hype, just signal.
Today's stories expose four critical layers: IPO pricing, agentic creative tools, enterprise access limits, and open-source home control.
HOW TO READ THIS Follow the estimated $70–$78 share-price range to the up-to-$1.404T fully diluted valuation at its upper end. The Verge says this could rank among the biggest IPOs; the valuation is not IPO proceeds.
Anthropic set an estimated IPO price range of $70 to $78 per share, a band that would value the company at up to $1.404 trillion on a fully diluted basis if shares price at the top, according to TechCrunch and The Verge. That valuation would place it among the world’s most valuable public companies, a striking marker for a company that was a research lab best known for safety-focused alternatives to OpenAI only a few years ago. The story leads today because it crystallizes how much capital markets are now willing to bet on frontier AI labs directly, rather than through proxies like Nvidia or Microsoft.
The range appears in Anthropic’s updated IPO filing ahead of a planned Nasdaq listing under the ticker ANT before the end of November, with Goldman Sachs and Morgan Stanley serving as underwriters. Price bands like this are set from institutional investor feedback during the roadshow, not from revenue multiples alone, so the final print can move meaningfully once order books close. TechCrunch reports that Anthropic disclosed an $81 billion revenue run rate by the end of October, $17 billion in third-quarter revenue and $33.2 billion in nine-month revenue.
What's novel isn't the IPO itself, which was already expected, but the specificity of the band and the scale it implies — a valuation in the same tier as the most valuable public companies on earth. A listing near the top of this range would hand Anthropic a large war chest and public-currency stock to use for acquisitions and compute deals, a potential durable edge over privately held rivals still raising at fixed valuations. The limitation is that price ranges routinely shift before pricing night, and a trillion-dollar-plus valuation invites scrutiny of whether growth can justify it once quarterly earnings become mandatory.
HOW TO READ THIS Adobe previewed a local MCP: an outside agent would connect to Creative Cloud apps on the user's computer and edit native project files, illustrated by a selected layer. Layers remain editable. This is a workflow schematic, not a working-product capture. MCP private beta is planned for early 2027; Builder is already in private beta.
Adobe used its MAX event in Miami to preview AI agents built into Photoshop, Premiere, Illustrator and InDesign, plus a local Model Context Protocol integration that would let outside assistants from OpenAI, Google and Anthropic act inside its apps. This is relevant because Adobe is previewing a bring-your-own-agent model rather than limiting users to a single in-house assistant.
The local MCP integration would expose app functions as callable tools so third-party agents could, in principle, edit a file in Photoshop or run tasks in Premiere the way coding agents already call external tools, mirroring how Anthropic's MCP standard works elsewhere. Firefly is also getting a prototyping tool called Builder, which TechCrunch frames as Adobe moving into Figma's territory, plus an 'always-on' agent mode for recurring background tasks. Builder is in private beta, while Adobe says the MCP integration enters private beta in early 2027; a VP told TechCrunch the company is still deciding what capabilities to expose.
The genuine novelty is less the individual features — object removal, relighting, blemish correction are incremental — than the architectural bet that creative work will be mediated by interchangeable third-party agents rather than a single Adobe assistant. The potential advantage is owning the surface area those agents need to touch, rather than competing model-for-model with OpenAI or Anthropic. The clearest limitation is timing: the MCP integration is slated for private beta in early 2027, while Builder is in private beta.
HOW TO READ THIS The gate depicts role-based coding-tool access: engineers and strats kept access; most traders and bankers lost it. The bottom bar separately shows approximately 12,000 engineers among roughly 47,000 staff, not an exact count of everyone retaining access.
Goldman Sachs updated internal guidance in October restricting Anthropic's Claude Code and GitHub's coding tools to its engineers and strats, cutting off access for most traders and bankers, according to Bloomberg reporting relayed by State of AI. It's worth flagging because it cuts against the assumption that enterprise AI agent adoption only ever widens, and because Goldman is simultaneously an underwriter on Anthropic's own IPO.
The restriction followed an internal review of productivity and risk concerns; traders had reportedly been using the tools to monitor client flows, model positions, and assemble written and graphic material, uses the bank apparently decided warranted tighter control once confined outside its engineering org. It's the second such pullback this year: Goldman removed Claude from its Hong Kong banker platform in April, reading its Anthropic contract strictly, with JPMorgan following suit on Claude access in Hong Kong in June. Goldman's own multi-model in-house assistant remains available firm-wide, and Anthropic engineers reportedly spent six months at the bank earlier this year building Claude-based agents now run by Goldman's own teams.
What's notable isn't that a bank imposed an AI restriction — finance has done that before — but that this one narrows access inside a bank that has otherwise been a visible, cooperative Anthropic partner, suggesting internal risk reviews can outrun partnership incentives. For vendors, it's a reminder that selling AI tools into regulated finance means budgeting for access to shrink as well as grow. The account rests on sourcing from people familiar with the matter rather than an official Goldman policy document, so the exact scope and permanence of the restriction aren't independently confirmed beyond that reporting.
HOW TO READ THIS Read downward from homeassistant-ai’s ha-mcp through assistant tool calls into dashboard and automation edits, ending with an example sunset rule switching on a porch light.
The open-source ha-mcp project under the homeassistant-ai GitHub organization is a Model Context Protocol server that lets Claude, ChatGPT and other MCP-compatible assistants configure, automate and control a Home Assistant smart-home setup directly. It made today's roundup because it reached #7 on GitHub's daily trending list, a sign that handing a general-purpose AI assistant control of home automation — rather than relying on Home Assistant's built-in rules engine — is resonating past a niche audience.
The project exposes Home Assistant's device and automation APIs as MCP tools, so an assistant can create automations, edit dashboards and flip devices through natural-language requests instead of writing YAML, and it installs via Home Assistant's HACS store, as a native add-on, through Docker, or via PyPI. The November 10 repository snapshot showed 6.4k stars and 365 forks under an MIT license, though the project's own documentation flags that its stdio connection method has known reliability problems and should be treated as demo-only for now.
Pointing an LLM at home-automation APIs isn't new, but packaging it as a standardized MCP server that works across Claude Desktop, Claude.ai and ChatGPT alike, rather than a one-off integration tied to a single assistant, is what's distinct here, and it previews a pattern likely to repeat across other control-plane software. Its GitHub following could help attract contributors, but Home Assistant already has other AI-assistant integrations, and the project's stdio transport has known connection problems. For a real setup, its documentation recommends the custom component or an HTTP method instead.
An agentic skills framework and development methodology for giving coding agents durable, reusable skills instead of re-deriving process every session.
Open-source terminal agent for enterprise Code Assist workflows and legacy paid API/Vertex keys; free/Google AI plans moved to Antigravity CLI, and it is not optimized for Gemini 4+.
An open-source agentic video production system with 12 pipelines and 700+ agent-skill and production-knowledge files, aiming to turn a coding assistant into a full video studio.
A free, benchmarked open-source memory system for AI agents, addressing the recurring problem of agents losing context across sessions.
An agent skill that researches a topic across Reddit, X, YouTube, HN and Polymarket and synthesizes a grounded summary; X and YouTube require source setup.
Ars Technica AI Ars Technica talks to mathematician Steven Strogatz about what happens to math once AI becomes something closer to an oracle than a tool.