Welcome back! Claude now has a seat in your Slack, a coworker you tag for tasks that already writes most of its own team's code. Today also brings biology you can program like software, Meta hitting pause on the AI that spied on its own staff and Mistral out-reading the field on documents.
In today's Generative AI Newsletter:
Claude: What happens when you can tag Claude in Slack like a coworker?
Biology: What does it mean that you can now code DNA like software?
Meta: Why did Meta pull the plug on its own employee-tracking AI?
Mistral: Why is Europe suddenly winning document AI?
Claude now has a seat in your Slack. Anthropic's new Claude Tag lets anyone in a channel tag @Claude, hand off a task and get back to work while it grinds through the thread.
Give it the channels, tools, data and codebases you choose, and it holds the context and plans ahead.
Anthropic has run its own product team on the tool all year, and says it now writes 65% of their code, including most of itself. Read that twice. The coworker that builds the coworker.
Then there's access.
Claude joins under one company identity with the keys to whatever an admin connects, so its reach is whatever an admin wires up, whoever does the tagging.
You get per-channel limits, spend caps and an audit log of every request, and you will want all of it before you plug an agent into your tools and source code.
It's in beta on Team and Enterprise, and it kills the old Claude in Slack app on August 3. Useful, obviously. An agent this capable is also only as safe as the access you hand it.
Ten years ago Nvidia was a chip company investors scrolled past. That $1,000 stake is worth about $225,000 now, pulled up by the AI boom.
The money has followed. In 2025 close to one of every two venture dollars went into AI. Khosla Ventures saw it early too, putting $50M into OpenAI in 2019 for a stake now worth billions.
In AI, the biggest outcomes have gone to people who got in while the company was still small.
GenAI Works is at that stage now.
Revenue almost tripled in a year running go-to-market for brands like Nvidia, Oracle, Google and IBM, and its enterprise platform ToneUp turns that engine into a product.
You can back it from $1,000. Sign by July 1, 2026 to lock Phase 1 bonus shares, from 5% at $1,000 up to 25% at $25,000.
7 days left.
In making an investment decision, investors must rely on their own examination of the issuer and the terms of the offering, including the merits and risks involved. Genai Works, Inc. has filed a Form C with the Securities and Exchange Commission in connection with its offering, a copy of which may be obtained here.
Biology is starting to look like a software stack.
The Arc Institute, a nonprofit research lab, released Proto, a programming language for designing living systems. You write protein, DNA and RNA the way you write code, or you describe what you want and let an AI agent compose it.
It comes from the lab behind Evo, the model that learned to read and write DNA across the tree of life.
The ambition is biology as a programmable layer, where you compose cells the way developers compose apps. Push it far enough and you are coding the human genome itself.
Temper the awe. This is a preprint rather than a product, and writing a working gene is nothing like shipping a web page, because biology fights back in ways code never does.
Still, the people who taught machines to read DNA now want to hand everyone a pen, and that is how a field gets rewritten, slowly and then all at once.
Meta built a program called the Model Capability Initiative that records most US employees' keystrokes, mouse movements and the odd screenshot, then feeds the lot to its AI to learn how people actually work.
This week it hit pause because the surveillance leaked.
Documents reviewed by Reuters showed the recordings, private chats, performance notes and transcripts, sitting where any Meta employee could read any other.
A tool built to study how staff work could not keep what it captured away from the staff.
Meta says it has no sign the data was misused and will investigate, with no date to switch it back on.
The program was already loathed.
Staff had protested being mined by software built to learn from them, right as layoffs landed, and Meta had bolted on a button to stop the tracking for 30 minutes at a time.
A pause with no end date is how a program like this comes back without anyone announcing it.
Feed a chatbot a scanned invoice or a PDF full of tables and it reads it back to you with total confidence and half the numbers wrong.
Reading documents is the thing AI is worst at and businesses need most. Mistral's new OCR 4 is the first model that looks like it cracked it.
Instead of dumping a wall of text, it labels what it sees, a table here, a signature there, then scores its own confidence on every word, so you can catch its mistakes instead of trusting them blind.
In blind tests, human reviewers picked it over every rival around 72% of the time, across 170 languages.
The lab that pulled this off is French. Mistral went after the document grunt work that actually runs a business, the contracts and invoices nobody wants to read, and made the best tool for it.
You can self-host it or pay about $4 per thousand pages. Unsexy, and exactly the kind of thing that wins.
Devin is an AI software engineer that takes a coding task and runs it end to end. You hand it a bug, a migration or a feature, and it writes the code, opens a pull request, fixes the failing tests and comes back when it's done. Nubank pointed a fleet of Devins at an 8-year-old codebase and says it cut a multi-year migration to weeks.
Try this yourself:
Sign up at devin.ai and connect a GitHub repo so Devin can see your code.
Give it a contained job, like fixing a flaky test or writing docs for a legacy file, and let it open the pull request.
Tag Devin in Slack or drop it a Linear ticket, and it turns the thread into a PR.
Spin up a fleet of Devins on a repetitive migration across files or repos.
Who it's for: engineering teams drowning in migrations, bug queues and the boring half of the backlog.
Google Flow can scout film locations in Street View: Its AI agent now generates images and video grounded in real Google Maps Street View, so you can drop a scene into an actual street, US locations first.
Allbirds is now an AI company: The struggling shoe brand sold its footwear business for $39 million, became Smartbird, pivoted to AI infrastructure and watched its stock jump about 40%, because putting AI in the name still moves the market.
Nvidia is wiring Europe with 35 AI supercomputers: It announced a record 35 new AI machines in development across 23 countries, Europe's biggest one-year compute buildout.
OpenAI beefed up Codex on mobile: Its coding agent's app added goals, inline code review and worktrees, so you can steer multi-step jobs from your phone.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
🐝 Subscribe to our Atlas newsletter — trusted by 3M+ subscribers — to stay ahead of AI news across tech, education, and business.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🚀 Explore investment opportunities in the future of AI and join our community-backed growth journey.
