Welcome, humans.

xAI's new coding agent, Grok Build CLI, got run through a network sniffer this week, and the results are ugly if you've ever pointed it at a private repo.

Turns out the tool doesn't just read the files it actually needs. It ships your whole codebase, full commit history, and anything sitting in your .env file straight to a Google Cloud bucket, and it does this whether or not you flip the privacy toggle off.

The numbers make it worse: on a 12GB test repo, the actual AI conversation used 192KB of data. The quiet background upload was 5.1GB. That's like asking a coworker to grab one folder off your desk and watching them walk out with the entire filing cabinet.

Here’s what happened in AI today:

😼 Satya Nadella called out model-lab distillation rules.
📰 OpenAI and Anthropic warned Washington that Chinese companies are using distillation to clone advanced U.S. AI models.
📰 Anthropic began localizing Claude pricing in India, its biggest market outside the U.S.
🍪 Waze added Gemini-powered destination search, conversational road reports, and motorcycle routing to bring AI into your everyday drive.

…and a whole lot more that you can read about here.

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

Love robots? We just launched a robotics newsletter! Sign up for it here.

😺 Satya Nadella calls out AI’s model-cloning double standard

The AI labs that learned from the entire internet would now prefer that nobody learn too much from them.

Microsoft CEO Satya Nadella has entered the growing fight over model distillation, arguing that frontier labs cannot champion broad rights to train on public information, then impose tight restrictions when competitors use their model outputs as training material.

Here's what happened:

Nadella criticized the one-way flow of knowledge between creators, customers, and AI companies.
Distillation lets developers train a new model using answers generated by a more powerful one, often reproducing valuable capabilities much more cheaply.
OpenAI and Anthropic warned Washington that Chinese companies are using distillation to clone advanced U.S. models. They say foreign competitors have used the technique at enormous scale, potentially converting billions of dollars in American research into a shortcut for rival labs.
A separate Business Insider investigation found that distillation could threaten the profits supporting the frontier-model business altogether.

The theft allegations deserve to be taken seriously. Anthropic says Alibaba used roughly 25,000 fraudulent accounts to collect nearly 29 million Claude interactions. That looks less like a curious researcher testing an idea and more like someone backing a data truck up to the loading dock.

But as we argued in our own report this morning, stopping illicit extraction and supporting open AI development are not mutually exclusive positions.

Why this matters: Frontier labs have spent staggering sums hiring researchers, buying computing power, and producing the datasets behind their best models. If another company can reproduce those capabilities for pennies on the dollar, the economic case for spending billions on the next model becomes much harder to defend.

The problem is that these same labs built their systems using books, articles, code, images, and other public material they frequently did not license individually. Their argument effectively becomes: learning from other people’s work drives innovation, while learning from ours threatens innovation. It’s simultaneously hypocritical and true.

Our take: Nadella has identified the principle the industry still needs to settle: who gets to learn from whom, under what conditions, and with what compensation?

Fraudulent accounts and industrial-scale extraction can reasonably cross a line. Broad restrictions on distillation could also lock meaningful AI research inside the handful of companies wealthy enough to build frontier models.

The labs may win stronger protections. They will have a much harder time convincing everyone that intelligence should flow freely toward them, then stop at their gates.

FROM OUR PARTNERS

How Visa triages alerts in seconds with agentic security — without losing the human touch

Visa used Elastic to build its first agentic AI workflow, automating a four-stage investigation pipeline with human oversight. Alert triage dropped from 20 minutes to seconds, producing fully enriched, incident-ready cases instantly.

See how Visa is bringing agentic AI into the SOC.

🎓 AI Skill of the Day: Run a Three-Line AI Cost Audit

Before you switch to the newest model, make the AI prove it deserves the upgrade. Today's pricing stories all point to the same reader problem: the best AI tool is not always the most expensive one, and the cheapest one is not always cheaper once you count failed attempts.

Use a three-line audit before picking a model or plan: task value, failure cost, and required quality. If the task is low-stakes, route it to the cheaper model. If a mistake creates legal, customer, or strategy risk, use the stronger model and ask it to show uncertainty.

I need to choose the right AI model or tool for this task: [describe task]. Classify it by task value, failure cost, required quality, and whether speed or accuracy matters more. Then recommend the cheapest safe option and explain what would justify upgrading.

Favorite insight: model choice is becoming budget management, not vibes. Want more tips like this? Check out our AI Skill of the Day Digest for July.

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!

Waze added Gemini-powered destination search, conversational road reports, motorcycle routing, and personalized navigation, bringing AI assistance into your everyday drive —free app.
Claude Fable 5 stayed available at no extra cost for paid Claude users through July 19, giving subscribers a longer window to test Anthropic's higher-end model —paid Claude plan required.
UnitPay handles all the billing infrastructure for your AI product, letting you launch usage-based, credit, or outcome pricing without writing a single line of code, and go live in 10 minutes.
Fabraix Playground lets you test your prompt injection skills against real AI agents, so you can see exactly how they break before an actual attacker does.
Osaurus turns your Mac into a home base for AI agents that remember, automate, and run entirely offline, so nothing you type ever leaves your machine.
AgentKey gives your AI agent one install for live web search, X, Reddit, YouTube, and on-chain data, so it can actually answer questions instead of hitting a paywall.

New from The Neuron: AI Explained

AI agents are moving beyond laptops into trucks, field teams, dispatch centers, and the real world.

Samsara CTO John Bicket explains what that actually looks like.

Recorded on-site at Samsara Beyond 2026 in Las Vegas, Corey and Grant sit down with John Bicket, Cofounder & CTO of Samsara, to talk about how AI is changing physical operations.
Samsara works with companies that run large fleets, field teams, drivers, technicians, vehicles, and physical assets. In this episode, John explains how Samsara is using AI across dash cams, telematics, asset trackers, Agent Studio, safety workflows, fleet data, and AI ride-alongs.

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

📰 Around the Horn

Google Research just gave LLMs something suspiciously human: bedtime. Right now, most models are stuck in an eternal cram session: great at short-term recall, but they lose it all once the context window closes. This "sleep" phase is basically Google's attempt to give models a nightly memory transfer instead of forcing a full retrain every time they learn something new.

OpenAI and Anthropic warned Washington that Chinese companies are using distillation to clone advanced U.S. models.
Anthropic began localizing Claude pricing in India, its largest market outside the U.S.
TechCrunch framed AI murder-planning questions as a live test of model safety thresholds.
More than 200 economists and AI leaders, including 16 Nobel laureates, urged faster preparation for AI's labor-market impact.
Meta expanded its Northeast Louisiana data-center commitment from $27B to more than $50B and five gigawatts of capacity.
Hacker News surfaced Grok CLI reports that turned local-agent file permissions into a security concern.
Axios reported OpenAI and Anthropic IPO wealth could reshape San Francisco housing demand.

Want absolutely EVERYTHING that happened in AI this week? Click here!

🔧 Tuesday Tool Tip: Ask for the Cheapest Safe Model

Next time you ask ChatGPT, Claude, or Gemini to do real work, add one line before you start: "Tell me the cheapest model or plan that can safely do this, and what would make you upgrade."

That forces the assistant to separate easy drafting, medium-risk analysis, and high-stakes work before you burn premium credits. Very glamorous. Extremely useful.