Welcome, humans.

The New York Times reported Wednesday that Meta labeled its AI data centers as experimental facilities so it could claim federal research tax credits. According to Quartz's recap, those credits shaved $3.9B off Meta's 2025 tax bill, up from $700M in 2023.

Meta's own accountants reportedly warned that the approach was legally risky. Meta went ahead anyway, which tells you how the math looked next to $3.9B.

Most experiments end with a lab report and a mess to clean up. This one ended with a $3.9B discount. Somewhere, a seventh grader with a baking soda volcano is asking about the paperwork.

Here’s what happened in AI today:

• 😺 Trump and top AI CEOs signed a voluntary safety accord

• 📰 OpenAI linked a data-extraction campaign to Moonshot AI

• 📰 Anthropic warned a Chinese open model can build cyberattacks

• 🍪 DoorDash launched ordering by text

• 🎓 Strip personal details before pasting anything into AI

JOIN US LIVE, LATER TODAY @ 10 am PT | 1 pm ET: We’re hanging out to break down everything that happened in AI so far this week. Come say hi and learn with us here

😺 Trump and the AI CEOs Signed a Voluntary Safety Pact (and Renamed AI "Super Intelligence")

Tuesday at the White House brought a rebrand and a pinky promise.

First, the rebrand: President Trump ordered the federal government to say "Super Intelligence" instead of "AI." Then, the pinky promise: the CEOs building the technology pledged to keep it safe.

Here's what happened:

• The rename: The executive order covers official federal communications only, leaves existing rules and contracts alone, and gives the president's science adviser 60 days to propose a legal definition of "Super Intelligence."

• The accord: Elon Musk, Mark Zuckerberg, Dario Amodei, Jensen Huang, Sundar Pichai, and OpenAI's Greg Brockman signed a one-page pledge. It promises four layers of safety checks, including internal testing, an outside audit, and a review by each company's board.

• The fine print: It's voluntary. Trump called it morally binding, but the accord only says that turning it into law may make sense down the road.

• The next-day twist: The FTC (the Federal Trade Commission, the government's consumer watchdog) confirmed Wednesday that it's investigating OpenAI, Anthropic, and others over AI's risks to consumers. The probe began this summer.

Why this matters: The backdrop is a rough few months. OpenAI and Anthropic have each reported cases of their AI agents slipping out of testing environments and carrying out cyberattacks.

For you, the practical read is simple. If you use ChatGPT or Claude at work, the protections in place right now are company promises plus one federal investigation. Sen. Mark Warner wants Congress to require testing, evaluations, and incident reporting for the most powerful models instead.

What to watch next:

• Trump said he'll name an "AI czar" (a single point person for AI policy) within days.

• The group discussed a 10-person committee to oversee the effort.

• The science adviser's 60-day deadline lands in late November.

Our take: Nothing says Super Intelligence like grading your own homework. The pledge leans on the companies' own checks, so the real test is whether that outside audit has teeth. The FTC, by contrast, started before any pledge existed and is reportedly drafting orders to compel executives to testify. Open question: if a lab breaks the pledge, who finds out, and what happens next? Watch who gets named czar this week; that pick will show whether this is the start of rules or the whole plan.

FROM OUR PARTNERS

The Enterprise Guide to Scalable AI

Plenty of companies can launch an AI pilot. Far fewer know how to turn that pilot into something secure, scalable, and useful in everyday work.

Explore “The Enterprise Guide to Scalable AI,” sponsored by Dell Technologies and NVIDIA. The hub covers what changes when AI moves from pilots to production, including how teams prepare data, choose infrastructure, run agents closer to users, and keep AI systems governed as they scale.

Learn More.

🎓 AI Skill of the Day: Strip the who, keep the what

Drawing a black box over text in a PDF hides it from you, not from a chatbot. A black box is a costume, not a lock.

Arpit Tripathi's privacy guide explains why: the box only covers the picture, while the actual characters (the "text layer") stay in the file, and an AI reads them right back. The safer move is deleting identifying details before you paste. The AI can still do the job, because it reasons about the situation, not the person.

Tripathi's method:

1. Decide what the task needs (the facts, not the identities).

2. Retype those facts into a blank note, swapping names for placeholders like [CLIENT], ID numbers for [NUMBER], and birth dates for an age.

3. Keep each placeholder consistent, so the AI doesn't lose track of who's who.

4. Paste your cleaned-up version with the prompt below, then fill in the real details yourself.

Test any "redacted" file: select all, copy, and paste into a blank note. If text appears, it isn't redacted.

I've replaced all names and ID numbers with placeholders in square brackets, like [CLIENT] and [CLAIM_NUMBER]. Treat each one as a stand-in and keep it exactly as written in your answer; I'll swap in the real details myself. Here's the situation: [paste your cleaned-up text]. Task: [draft the letter / summarize this / find the problems].

Passwords and API keys (codes apps use to talk to each other) never go in, placeholder or not.

Want more tips like this? Check out our AI Skill of the Day Digest for October (link pending).

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

FROM OUR PARTNERS

31 Days of Answers on Running AI Agents Securely

How do you give every agent its own identity? What happens when one goes rogue mid-task?

Google Cloud's Advent of Agents answers one question a day through October, each with a short practitioner video and code you can build with your coding agent. No sign-in, no paywall.

Start Day 1

🍪 Treats to Try

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

1. *Build your own AI coworker in just 3 hours (for $0). Join Outskill’s live workshop to learn the ins & outs of the 15 most powerful AI tools right now.

2. DoorDash builds your cart and checks out inside a text thread when you message it what you're craving or tell it to reorder your usual (US beta, waitlist open).

3. VoiceCap joins your Zoom, Meet, and Teams calls (or records the room) and turns them into searchable transcripts, summaries, and action items in 100+ languages —free for 300 minutes, then €29/month.

4. PixelCrew turns a written design brief into a finished landing page or dashboard in 25 to 45 minutes, using software agents that handle research, design, and QA —free in alpha with your own API key.

5. WZRD converts a PDF, PowerPoint, or spreadsheet (or just a prompt) into an interactive deck, doc, sheet, or form with a built-in voice agent that talks visitors through it and captures their answers.

6. Plane Agents are ready-made teammates inside the project-management tool Plane that compile your standup update, flag work at risk of slipping, and sort incoming requests, either on a schedule or whenever a task changes.

7. SereneDB is a database that finds records by keyword or meaning and crunches fresh numbers in the same query, using SQL (the standard language for asking databases questions).

📰 Around the Horn

• OpenAI said a core group behind a summer campaign to copy how its models reason is linked to Moonshot AI.

• Synopsys struck a revenue-share deal with OpenAI to build GPT-Synopsys, a model that helps engineers design computer chips.

• Anthropic said Z.ai's open GLM-5.3 model can build working cyberattacks and that simple tricks got around its safeguards in tests.

• Google said publicly reported software security holes doubled to 10,740 in August, as hackers use AI to turn newly patched flaws into attacks.

• DeepMind launched SynthID Bio, a watermark for AI-designed proteins that didn't hurt how well they worked in lab tests.

🧩 Thursday Trivia

You know the drill. Which is real? Which is AI?

A

B

🎙️ New from The Neuron: Voice AI’s Biggest Blind Spot

AI voices pretty much sound human now. So the real challenge is whether they actually understand what we humans mean when we speak.

We sat down with Andrew Ettinger, CEO of Hume AI, to dig into voice AI’s “I’m fine” problem (what does it mean?): because voice AI mostly reads transcripts, they capture your words but misses your tone, emotion, pauses, accents, noise, and all the other subtle signals that tell you what someone actually means.

Andrew also explains how Hume evaluates voice models and why the winner in voice AI won’t just be the one that sounds most human for 30 seconds.

Watch and/or Listen now: YouTube | Spotify | Apple Podcasts

A Cat’s Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S. Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.