Welcome back! Google's new $99 Fitbit Air comes with an AI health coach that will tell you to skip your workout, hydrate more, and eat bland foods when you're exhausted. The catch? It only works well if you spend six hours teaching it your medical history. Oh, and don’t forget to manually type in blood test results and remind it of your previous conversations.
Google made the AI coach completely optional, so you can ignore it entirely and just use the tracker. Is opting in to AI health advice, rather than forcing the issue, the right approach? Hit reply and let me know.
What If You Could Teach AI by Just Doing the Task Once?
OpenAI just rolled out Record & Replay for Codex, and it's one of those quietly huge updates that I think a lot of people are going to be talking about soon.
The gist: you hit record, do a task once on your Mac, and Codex watches. It turns what you did into a reusable skill. The next time you want to do that same task, you just call the skill and Codex handles the whole thing.
A few things stand out:
It captures intent, not pixel coordinates. Old-school RPA tools like UiPath break when an interface shifts by a single pixel. Codex generates a natural-language SKILL.md file that its model interprets against whatever the screen actually looks like at the moment.
You can open and edit that skill file. It's not a black-box recording—you can read what Codex learned and tweak any step that got captured wrong.
It works across browser, computer use, and connected plugins. Slack, Gmail, Notion, etc. or any combination.
My POV: The fiddly micro-decisions that make a workflow yours (which spreadsheet column the title comes from, which folder you save to, which dropdown option you always pick) are nearly impossible to put into a prompt. But they survive a recording perfectly.
Pair this with the Codex on Mobile update from a couple weeks ago, and you've got a setup where you teach Codex something on your Mac once, then kick it off from your phone while you're across town. That is the agentic AI workflow a lot of us have been promised but haven't actually felt yet.
A few caveats: it's macOS only at launch, not available in the EU/UK/Switzerland, and you need an active ChatGPT subscription with Computer Use enabled. But assuming it expands, this is going to change how a lot of people work.
— Matt
Samsung Gives ChatGPT & Codex to Employees Worldwide
Via The Independent
Samsung Electronics is rolling out ChatGPT Enterprise and Codex to its employees in South Korea, plus all Device eXperience employees worldwide.
The rollout: OpenAI called it one of its largest enterprise deployments to date. Samsung plans to use ChatGPT and Codex across R&D, software development, product development, marketing, manufacturing, corporate functions, and other parts of the business.
Why Codex matters:
Codex is still being used for developer work like writing, reviewing, and debugging code, but OpenAI says it is also becoming useful for non-technical teams. Employees can use it to turn ideas into internal tools, websites, automated workflows, and working software.
More than 5 million people now use Codex every week. In Korea, Codex weekly active users have grown nearly 800% since the beginning of February.
The bigger picture: This is what enterprise AI looks like after the pilot phase. Samsung isn’t just giving AI to one small group of developers. It’s pushing ChatGPT and Codex across technical and non-technical roles, which is how AI actually becomes normal at work—not as one magical app, but as a default layer across teams and functions.
Microsoft Taps Chevron to Power AI Data Centers
Chevron signed a 20-year deal with Microsoft to build a natural gas power facility for a new data center campus in Pecos, Texas.
The deal: The project, called Kilby, will provide dedicated electricity to Microsoft’s West Texas data center campus. Microsoft says the campus will add 2 gigawatts of data center capacity, while Chevron says the power project could eventually ramp up to 2.67 gigawatts. The facility is expected to start delivering power in 2028.
The buildout: Most of the generation will come from GE Vernova turbines, with additional capacity from Caterpillar subsidiary Solar Turbines. Microsoft says the larger data center investment will support more than 6,000 construction jobs and hundreds of permanent operational roles.
The bigger picture: The AI race is becoming an energy race. Models get the attention, but the bottlenecks are increasingly physical: power, land, cooling, chips, data centers, and grid access. Microsoft’s 20-year power deal with Chevron is a pretty clear signal that the next phase of AI will be shaped as much by infrastructure as software.
Stop Renting Chatbots. Hire an AI Employee.
You don't need another tab to prompt. You need someone to hand the work to.
Viktor is an AI employee who lives in Slack and Teams, connects to 3,200+ tools, and actually does the job. Not "here's a draft." Done.
Put him to work, and he'll:
→ Reconcile the books and chase the overdue invoice before you've finished your coffee
→ Audit the ad account, flag what's bleeding, and write the fix
→ Pull the numbers, build the report, and post it in the channel
→ Onboard the new hire while you're in your 1:1s
One hire. Your whole team suddenly operating like it's twice the size.
A visual dictionary for better AI image prompts
Via AtomWords
AtomWords is a free visual dictionary of AI art keywords that helps you find the right words for aesthetics, mediums, materials, colors, moods, lighting, composition, and camera styles. Instead of guessing how to describe a look, you can browse prompt-friendly terms like golden hour glow, bioluminescent lighting, or impasto oil painting.
How you can use it
Find better words for image prompts
Explore visual styles before generating
Build more consistent aesthetics across images
Discover lighting, material, camera, and mood terms
Pricing: Free
Open-source AI agent workflows
Via AgentDock
AgentDock is an open-source platform for creating and deploying AI agents and automated workflows. It includes a visual workflow builder, modular architecture, third-party integrations, and enterprise-grade security controls while staying framework-agnostic so teams are not locked into one model provider.
How you can use it:
Build internal AI agents without starting from scratch
Create automated workflows across business tools
Prototype agentic systems with a visual builder
Connect third-party integrations into one workflow
Pricing: Free
Jobs, announcements, and big ideas
ByteDance unveils Seedance 2.5, a model that spins 30-second videos out of a single text prompt.
Meta partners with EssilorLuxottica to launch a new line of consumer-ready Meta Glasses.
Anthropic introduces Claude Tag, a new way to label and organize work across Claude.
Getty Images’ makes a deal with OpenAI for licensed content libraries to appear across search and discovery within ChatGPT.
NVIDIA launches its BioNeMo Agent Toolkit, arming AI agents with tools to accelerate scientific discovery.
OpenAI debuts Patch the Planet, helping open-source maintainers find and fix vulnerabilities.
You thought OpenClaw was insane? You haven’t tried Odysseus yet. Watch along as I break it down.
That’s a wrap! See you Friday for more.
