May 21, 2026 | Listen online | Read online

Welcome, humans.

Voice agents are having their moment. Every demo looks clean, every pitch sounds effortless, and everyone is very confident their AI will just...work. So naturally, we decided to do a live demo.

This Thursday we're hosting a livestream with Ben Cherry of LiveKit, the open-source platform behind a lot of the voice, video, and physical AI agents actually shipping in production today.

We're going deep on what it really takes to build a voice agent that can listen, respond, interrupt, call tools, and not fall apart the moment a real user says something unexpected. Then we'll build one live, on stream, together. There is at least a small chance the agent talks back at exactly the wrong time. We're told this is a feature.

See you at 10AM PT | 12PM CT | 1PM ET on YouTube, LinkedIn or Studio Link.

Here’s what happened in AI today:

😹 Meta used its own employees' keystrokes to train AI, then laid off 8,000 of them.
📰 OpenAI is filing for its IPO, targeting a September debut at an $852B valuation.
📰 Grok launched Skills: teach it something once, it remembers forever.
📰 The White House is requiring AI companies to share new models with the government 90 days before launch.
📰 Airbnb added groceries, car hire, and boutique hotels. It's basically becoming a full travel OS.

Hey: Want to reach 700,000+ AI-hungry readers? Advertise with us!

P.S: Love robots? We’re starting a new robotics newsletter! Sign up early here.

😺 Meta Used Its Employees as AI Training Data, Then Laid Them Off

A leaked audio recording reveals Zuckerberg's real explanation for why Meta was tracking its workers.

You've heard of "learn by doing." Meta invented "learn by watching employees do it, then replace them."

A leaked audio recording from a Meta all-hands meeting, obtained by More Perfect Union, captures Mark Zuckerberg explaining that Meta has been monitoring employee activity across Gmail, GChat, internal tool Metamate, and VSCode (the coding software most engineers use) to train its AI models. His reasoning: the AI "learns from watching really smart people do things," and elite engineers make better training subjects than outside contractors.

Which is, in fairness, a solid training strategy. It's the "then lay them off" part that makes it weird.

Here's how this unfolded:

April 21: Meta announced that it was installing keystroke and mouse-tracking software on employee computers. Meta's public response was mild: the models just needed to learn how people click dropdown menus. Routine stuff.
April 30: At an internal all-hands, Zuckerberg gave a rather different explanation on tape. The AI learns from watching "really smart people." He also acknowledged the leak risk, telling staff it was "not strategically in your interest" to share details openly.
May 19 (Monday): Meta announced it was reassigning 7,000 workers to new AI-focused teams. Framed internally as a productivity upgrade.
May 20 (Wednesday): Roughly 8,000 employees received layoff notices, starting at 4am Singapore time. The leaked audio dropped the same day. Fliers opposing the tracking program appeared on office walls.

Why this matters: The gap between the public story ("we're just training models to use software") and the private one ("we're learning from watching our best people") is the whole story. Meta is hardly alone in collecting employee behavior data. It's just the company that got caught explaining the real logic on tape, right before laying off the people it was watching.

Our take: Meta will survive this. The more uncomfortable question is whether this becomes a template. Every company with a "productivity monitoring" program is now one leaked memo away from the same headline. The line between "helping you work better" and "training your replacement" just got a lot blurrier.

FROM OUR PARTNERS

Ready to move past novelty chatbots? Google Cloud’s Startup School: Agentic AI covers the architecture you need to build robust, autonomous workflows.

Dive into Realtime Voice AI, enhanced Multimodal RAG, and bidirectional Vision Agents. Go from prototype to secure production without missing a beat.

Save your spot

🎓 AI Skill of the Day: Turn Claude Into Your Full-Time Content Team

If you make content for a living (or want to), Modern Millie's 23-min tutorial is worth every minute. Here are the four workflows that actually move the needle:

1. Free profile audit. Install the Claude Chrome extension, open your Instagram or YouTube Studio analytics, then ask: "Audit my profile. What's working, what isn't, and what's the single most important thing I should change right now?" Claude reads the live page and gives you a full breakdown — hook strength, content gaps, growth opportunities.

2. 30-day content calendar in one prompt. Inside a Claude Project loaded with your brand info, paste this:

Based on everything you know about my brand and goals, 
build me a 30-day content calendar. Alternate between 
my content pillars and include 3 hook options per video.

3. Full video script from the calendar. Pick a day you like, then:

Let's work on Day [X]. Use Hook [#]. 
Turn this into a full script with text hook, verbal hook, 
visual hook, talking points, and CTA.

4. One video → one week of content. Get your YouTube transcript from transcript.io, start a new Project called "[Brand] YouTube to Everywhere," and paste: "Here's my transcript. Turn this into a full week of content." Out comes short-form scripts, captions, carousel outlines, and an email newsletter — all from the video you already made.

Total AI beginner? Start here (goes with this video).

Have a specific skill you want to learn? Request it here.

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

The World’s first Claude-a-thon, A 2 day deep dive into Claude, its user-cases and 10+ other AI tools happening this weekend from 10 am-7pm EST. Register NOW! (free for the next 48 hrs)
Dell Pro Max with GB10 helps teams turn AI ideas into pilots, demos, and workflows with NVIDIA Grace Blackwell power and 128GB memory. See it here.
PollyReach gives your agent a real phone number and sends it out to make calls (qualifying leads, booking appointments, or handling customer support) while you do literally anything else.
LandingHero AI adds a voice agent to your website that talks to visitors, answers product questions in 50+ languages, and captures leads with full transcripts; add it to any existing site with one line of code.
LobeHub manages your entire agent team for you: assign a goal, it picks the right agents, runs them in parallel 24/7, and only contacts you when a decision actually needs a human. —free to try
NanoClaw is a sandboxed alternative to OpenClaw (a popular open-source AI coding agent) that runs in a container so your AI agent can't access your entire machine. It went viral,, turned down a $20M buyout, and raised $12M seed with the Hugging Face CEO as an angel.
Lance is ByteDance's open-source model that handles image generation, editing, video generation, and video understanding all in one framework. At 3B parameters, it benchmarks ahead of most 7B models at a fraction of the size.

Featured: The AI Trying to Solve Math’s Biggest Mystery w/ Tudor Achim of Aristotle

What happens when AI stops simply giving answers and starts producing proofs a computer can verify?

In this episode of The Neuron, Corey Noles and Grant Harvey talk with Tudor Achim, Co-Founder and CEO of Harmonic, the company behind Aristotle — a formal reasoning system built to generate machine-checkable mathematical proofs.

Tudor explains why math may be the clearest test case for moving AI from “trust me” to “check me,” and why formal verification could matter far beyond Olympiad benchmarks.

Special thanks to our sponsor Dell & NVIDIA for making this one possible!

New episodes air every week on: Spotify | Apple Podcasts | YouTube

📰 Around the Horn

The White House briefed OpenAI, Anthropic, and Reflection AI on a planned executive order that would let government agencies review advanced AI models up to 90 days before public release; Trump could sign as early as today, with cloud providers, chip companies, and banks also in the room.
OpenAI is preparing to confidentially file for an IPO as soon as this week, targeting a September debut at an $852B valuation, with Goldman Sachs and Morgan Stanley leading the deal, one day after Elon Musk lost his lawsuit trying to block the company's restructuring.
SpaceX's public S-1 filing revealed Anthropic will pay SpaceX $1.25B per month through May 2029 for xAI data center access, and that X lost $595M in ad revenue in 2024 due to advertiser exits; IPO will trade as SPCX on Nasdaq.
Grok launched Skills, a persistent memory feature that lets you teach Grok a formatting rule, workflow, or personal preference once, and it remembers it across every future conversation, no re-prompting needed.
Google's Gemini became the #2 AI referral source in Q1 2026, nearly tripling its share to 11.6% and now sending more web traffic than Perplexity, Claude, Meta AI, DeepSeek, and Grok combined; ChatGPT fell for the first time.
Airbnb's summer release added car hire, grocery delivery, airport pickups, boutique hotels, and AI-powered review summaries to the app, expanding the platform well beyond home rentals for the second year running.

FROM OUR PARTNERS

Put the AI lab within reach

For business owners and AI enthusiasts who want more than browser-based tinkering, Dell Pro Max with GB10 brings NVIDIA GB10 Grace Blackwell power to a compact AI development desktop.

Check out the Dell Pro Max with GB10.