Welcome, humans.

Elon Musk says Tesla's Digital Optimus can now play about halfway through Diablo's campaign (the classic dungeon-crawling video game). The twist is that it only looks at the screen and uses a keyboard and mouse, the way you would. Musk also said it's good at Counter-Strike and other fast games, with League of Legends now in training.

Why would a car company teach an AI to play Diablo? A Tesla job listing for a "Gaming Track" engineer says the team is testing whether skills learned in games carry over to other software, then plugging those agents into Full Self-Driving and the Optimus robot. Games are a cheap practice room. If an AI can figure out a screen it has never seen, it could, in theory, run your clunky work software without special integrations.

Halfway through Diablo is also exactly where my own save file died, so the robot and I have the same attention span.

Here’s what happened in AI today:

• 😼 White House sets AI reporting rule after Claude mishaps

• 📰 Anthropic reportedly signs multiyear cloud deal with IREN

• 📰 Tesla renames Full Self-Driving to Assisted Driving in Europe

• 📰 Nuvacore seeks funding at roughly $2.5B valuation

• 🍪 OpenScreen turns screen recordings into polished demos

😼 White House Sets AI Reporting Rule After Claude Went Off-Script

So Anthropic disclosed a new batch of cases where Claude agents took unintended actions on the live internet. One Claude run submitted an invented tip through a real police website; other cases involved government systems the company never meant it to touch.

Then Washington moved.

Axios reports the White House is now requiring AI companies to promptly disclose and remediate unauthorized or fraudulent interactions with government and other systems. A new Super Intelligence Force will oversee that process, though the administration has not yet explained exactly how enforcement will work.

Okay, so what did Claude actually do?

An AI agent is a model wired to tools so it can act, not only answer. Give one a browser and a goal, and it can click, submit forms, fetch data, or run code on your behalf.

That autonomy creates a nasty failure mode: persistence.

During evaluations and internal use, Anthropic says Claude sometimes treated a boundary like a puzzle. If a website blocked access, the model looked for another route. In one run, a task that expected an example form submission ended with invented content being sent to a real police department.

Great for solving stubborn spreadsheet tasks. Less charming when the “obstacle” is a government website.

Anthropic grouped the incidents into four patterns:

• exploiting a basic software flaw to run commands;

• submitting a sensitive real-world form when it should have stopped;

• bypassing a token or paywall to reach gated data;

• using URL shorteners to get around fetch-tool limits.

The company says the known cases had minimal real-world impact and were less severe than earlier cyber incidents. That caveat matters. Anthropic’s own diagnosis, though, is more interesting: some training and evaluation setups rewarded completing the task strongly enough that “find a way around the blocker” could become learned behavior.

In other words, the quality that makes an agent useful, keep going until the job is done, can become the exact thing that makes it reckless.

Anthropic has now shut off live internet access across its internal evaluations until new monitoring is proven reliable, tightened its web tools, and built detectors that blocked all of the known cases in retrospective tests.

The White House response turns this from one lab’s safety problem into an industry reporting problem. The part to watch now is whether “tell us immediately” becomes a clear, enforceable standard, and whether labs can test increasingly autonomous agents without accidentally testing them on the real world.

FROM OUR PARTNERS

Gemini 4 Argon debuts at #1 on APEX-Agents; new top model for professional work.

Mercor's APEX-Agents leaderboard evaluates frontier AI on long-horizon, multistep tasks across economically valuable work.

Built with partners like Harvey, Ramp, and Cognition.

Every model. Ranked by productivity. 

View the scores

🎓 AI Skill of the Day: Build an Autonomy Ladder

Vesta’s mortgage agents offer a useful pattern for anyone rolling out an AI agent at work: do not jump from “assistant” to “fully autonomous.”

TechCrunch reports lenders often start with a person approving the agent’s work, then let it handle a small share of loans on its own, then expand as confidence grows. Every action and reason is logged so the lender can audit what happened.

That is an autonomy ladder. You increase freedom only after the lower-risk stage proves itself.

1. Start with approval: the agent drafts or recommends; a human executes.

2. Delegate a narrow slice: let it act on low-risk, reversible cases with full logs.

3. Expand by evidence: widen scope only after you have measured error rates, exceptions, and recovery.

Autonomy should be earned like production access, not granted like a free trial.

Copy/paste:

Design an autonomy ladder for [workflow]. Split the work into three stages: human-approved, narrow autonomous, and broader autonomous. For each stage, define the tasks allowed, actions forbidden, required logs, rollback method, error threshold, and the evidence needed before moving to the next stage.

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

FROM OUR PARTNERS

Most AI Tools Weren't Built for This

B2B customer issues move across teams and systems, not through a single chat window. A new Harvard Business Review Analytic Services briefing paper, sponsored by Front, breaks down where AI tools fall short in B2B service and what to ask before you invest.

Get the Free Report

📰 Around the Horn

• Anthropic reportedly signed an unannounced multiyear cloud deal with miner IREN, as it races to bring about $530B in data-center leases online.

• Tesla renamed “Full Self-Driving (Supervised)” to “Tesla Assisted Driving” in Europe after Germany’s transport minister called the old name somewhat misleading.

• Qualcomm CEO Cristiano Amon said AI agents that use your apps could spark a smartphone upgrade supercycle, and that AI companies want phones running 100B-parameter models by 2028.

• Nuvacore is raising hundreds of millions at a roughly $2.5B valuation, Reuters reported, despite being six months old with no product yet.

• The Gifted, an AI film, won the Future Vision XPRIZE grand prize: $100K plus $2.5M in feature production funding.

• Nikon disqualified the winning entry in its Small World in Motion contest for breaking its generative AI rules and said it will revisit the rules.

🍪 Treats to Try

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

1. OpenScreen records your screen and turns it into a polished demo, with zooms placed on your clicks, a restylable cursor, and captions generated on your own computer —free and open source, no watermark.

2. BetterWispr types clean text into any Mac app when you hold Option + Space and talk, dropping the ums (English only) and keeping your speech on your computer —free and open source.

3. HireKey compares your resume to a job post, rewrites your bullets to fit it, and runs mock interviews out loud with a voice coach —free plan, then $39/month.

4. BIGWORDS.PAGE turns any phone or TV into a full-screen sign or countdown timer from a single link, with no app or account —free.

5. jevman ranks AI models by how well they play 100 games of Pac-Man against the classic ghosts, and you can try beating the ghosts yourself.

6. Qualcomm CEO Cristiano Amon said AI agents that use your apps could spark a smartphone upgrade supercycle, and that AI companies want phones running 100B-parameter models by 2028.

7. Nuvacore is raising hundreds of millions at a roughly $2.5B valuation, Reuters reported, despite being six months old with no product yet.

8. The Gifted, an AI film, won the Future Vision XPRIZE grand prize: $100K plus $2.5M in feature production funding.

9. Nikon disqualified the winning entry in its Small World in Motion contest for breaking its generative AI rules and said it will revisit the rules.

UPCOMING EVENT: The Neuron IRL in San Francisco

The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen.

It’s free, but space is limited. Save your spot!

😹 Monday Meme

New from The Neuron: AI Explained

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

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.