In partnership with
Have you ever read an AI meeting summary and thought "wait, I didn't say that"?
In my recent conversation with Alex Himel, Meta's VP of Wearables, I asked him, "Why do I still have to say 'Hey Meta, take a picture'? If the glasses can see what I'm eating, why not just log the calories?"
If you want to know what his response was, click here:
To my surprise, the same week that video went out, Meta glasses owners on Reddit started finding conversations in their Meta AI app that they never asked the glasses to record.
One owner saw it on three different pairs, with "Respond without Hey Meta" switched off. Meta says the issue "has been resolved" and hasn't said what caused it.
The timing honestly couldn't have been better.
And well.. this isn't new.
Back in 2024, an AI researcher finished a Zoom call with some investors, and Otter emailed him the transcript, plus hours of the investors talking among themselves after he'd left.
More on that in a bit. First, here's what happened this week.
An OpenAI model spent money nobody gave it. Read more
OpenAI's latest misalignment report says an internal model was asked to deploy code on a crypto test network without being given a funded wallet.
So it checked 362 accounts derived from recovery phrases developers commonly reuse, found one with enough test tokens, paid with it, and left that out of its final report.
Character.AI filed for bankruptcy. Read more
The chatbot app that was valued at $1B in 2023 filed for Chapter 11 last Monday. Disney has offered $20M to license some of its IP.
Anthropic is hiring way fewer juniors. Read more
Its own numbers show junior technical hires dropping from 27% to 9% since early 2025.
Replit says it dropped Anthropic. Read more
Replit's CEO said onstage at TechCrunch Disrupt that it no longer works with Anthropic, saying "ChatGPT is better now."
Claude figured out in under 30 minutes where the NSA's mystery numbers in a 25-year-old encryption standard came from. → Read more
By August, agents at Anthropic were clocking 6x as many hours as its employees. → Read more
Anthropic wants to put 5% of its shares into a public benefit fund. → Read more
OpenCode's Console got breached, and some users say they're already being billed for usage they didn't run. → Read more
ChatGPT Slides is rolling out to Pro, Business and Enterprise users. → Read more
Most of us aren't wearing Meta glasses to work anytime soon. But most of us are on calls with a notetaker in them, and usually it's one somebody else set up.
That's pretty much what the Otter lawsuits are about too.
The people suing say they were recorded by someone else's Otter bot without agreeing to it, and in August a judge let their core privacy claims go ahead.
That's what happens when nobody agreed. Even when everyone did, there's a second problem: the summary doesn't always say what you said.
Nobody reads a 40-minute transcript, we all read the bullets. And bullets are short, so your "maybe" and your "only if" are usually the first things to go.
You say, "We could maybe look at this next quarter, if the budget frees up." The summary says, "Revisit next quarter."
I mean, where did the "maybe" go?
Lilach Bullock, who writes about AI tools for small businesses, had a notetaker stitch two sentences from different parts of a call into one neat statement.
When something goes wrong and someone asks "didn't you say Friday?", nobody opens the recording. They open the notes, so the AI's version of you is the one that sticks.
Ask before you talk. On any client or investor call, ask: "Are notes running, and who gets them?"
Give "maybe" somewhere to go. If you run the bot, set up a custom summary template with a section called "Open questions, undecided." Your maybes land there instead of turning into action items with dates on them.
Say your conditions out loud, as their own sentence. Something like: "This isn't a commitment yet. I'll send it Friday, only if the numbers come in."
Read your part of the summary the same day. If a line turns an idea into a promise, reply in the thread: "Small correction to the notes: I said I'd look into X if the numbers come in, not that I'd deliver it Friday."
MyLens AI Turns your material into one-on-one AI presentations.
Craft Agents A desktop app for working with AI agents that can run tasks on a schedule.
Fish Audio Drama 3 A voice model where you describe the emotion in plain words, and the tone can shift halfway through a line.
My Take
Every time you say "Hey Meta" or click to start the notes, you're giving consent. Every AI tool is now trying to remove that step, because not having to ask is the selling point.
This week OpenAI showed where that can lead: a model that needed money, couldn't get any, and decided on its own that someone else's wallet was fair game.
So the skill I'd build is deciding where your AI acts on its own and where it still has to ask, before the tools make that call for you.
Most teams haven't drawn that line yet, and the people who draw it first will be the ones others trust with these tools.
Would you let your AI notetaker join meetings without asking you first?
Hit reply and tell me.
Until next time,
Vaibhav 🤝🏻
If you read till here, you might find this interesting
#AD 1
Stop feeding the feed.
You’ve put in the work to build an audience. But when the only way to reach your fans is through a social feed determined by algorithms, the platform decides who sees what you make next.
beehiiv gives you the tools to reach your audience directly. On their all-in-one platform, you can turn followers into subscribers, grow your newsletter, and connect it all to your community or podcast. Stop chasing virality on platforms you can’t control. Build a stronger, more sustainable business on beehiiv.
#AD 2
Take the prompts our creative team actually uses
The key to maintaining brand quality at scale? Mastering how to communicate with AI models and embedding them into your creative strategy.
Join our honest discussion with Ari Murray, Chief Digital Officer at Salt and Stone, and get 5 tips for expert LLM prompting with ready to run prompts.
