Howdy wizards,
I've been a Granola user for a year, and recently they've added two super handy things:
1. Voice typing. You no longer need a separate subscription to a transcription tool like Wispr Flow.
2. Screen capture when someone shares a screen. Granola takes screenshots of their screen share (and nothing else around that) and adds them to your notes. Super helpful context later.
Shout out to Granola for shipping actually useful AI-features and for being a long-term sponsor of this publication.
Hereâs whatâs brewing in AI.
The big thing
Claude can now read your bank accounts, if you let it.
Claude has a Personal Finance page in the side menu now.
The way it works is you sign in to your bank through Plaid, a service that connects apps to banks. ChatGPT has had this since May.
The important parts: Claude doesn't see your password, and it can only read information â not take action.
It allows you to ask Claude about your money in any chat and get an informed answer.
Anthropic says there are no ads and that they have no financial products of their own. But if you wonder how they'll make money from it, there's a strong clue at the end of their launch post:
translation: we want to be a broker for your financial services
Image source: Anthropic's announcement
Anthropic says their personal finance app doesn't train on the personal finance data specifically. That might be true, though I'm wondering how they'd keep a clean separation when Claude can access the data from any chat...
Btw it's still in beta and only on Pro, Max and Ultra. Right now it's US-only.
Why it matters
You know how this will go:
2026: AI knows your finances
2027: AI pays on your behalf
Some of us are already giving credit cards to AI and letting it shop, but it will be much more commonplace soon.
PS if you want to track finances with AI, with full customisation freedom, building your own tracker has never been easier. I've built a personal finance dashboard for myself that I'm quite happy with:
My own vibe coded personal finance app. Data stored in Supabase, Claude has MCP access.
It's built to answer the exact questions I have, so I mostly don't need to prompt Claude to ask, although I can if I want.Â
IN PARTNERSHIP WITH GRANOLA
You know that feeling when you leave a meeting and immediately forget half of what you agreed to?
That's not a memory problem. When you're back-to-back all day, there's simply no time to process. No time to write the follow-up.
Granola helps you become the person who actually does what they said they'd do.
You take notes during the meeting. Just quick bullets, nothing formal. Granola transcribes in the background and turns those notes into clear summaries with actual next steps.
No more "wait, what did we decide?" moments. Just clarity. And follow-through.
NEWS NEWS NEWS ⊠NEWS NEWS NEWS
All the small things
Industry moves
Nvidia CEO Jensen Huang admitted an affair with a suspected Chinese spy. "I was unfaithful to my wife," he wrote to staff on Sunday. Apparently, the spy is a former executive at Megaspeed, a Singapore company that buys Nvidia's chips and that US authorities suspect of helping China get banned AI chips. Nvidia's board says it has "no reason to believe" any secrets were compromised.
US voters wanted limits on data centres in 46 out of 47 local votes last week. Meanwhile, OpenAI is working on their framing: they now want to rebrand their data centres to "AI compute campuses".
Research
METR spent 11 weeks inside Anthropic and says AI has roughly doubled the pace of model development. On Anthropic's internal tests, its models improved about twice as fast in the past year as before, and METR credits most of that speed-up to AI.
People follow AI advice even when it's random. In an Anthropic experiment, 50 analysts made judgement calls with an AI agent's help. A coin flip had secretly picked its opinions, yet they went along 86% of the time, and none noticed. In case you needed another great reason not to rely on AI for judgement and direction.
Talks & tutorials
Hadley Wickham checked whether AI is secretly sold below cost, and says it is for subscriptions but not for API usage. He worked out that running an open AI model on your own busy server costs about $0.25 per million tokens, while typical providers charge about $1.13 for that same model. So API prices cover the cost, with room to spare. Subscriptions are the exception: $200/month gets you â$2,000-10,000 worth of tokens, by some estimates. Long story short: subscriptions are cheap, API isn't.
âŠ
You are a delight.
What's your verdict on today's email?
Subscribe to participate
