• The Neuron
  • Posts
  • 😹 Grok killed a whole town in 4 days

😹 Grok killed a whole town in 4 days

PLUS: One company burned $500M on AI in a month

Author

Eric Gerard Ruiz
May 31, 2026

In partnership with

Somewhere in corporate America this week, a company burned $500 million in a single month on AI tools. Not because they were building something big. Because they forgot to set a spending cap on employee licenses.

The culprit has a name: "tokenmaxxing" (the corporate habit of maxing out AI usage without ever measuring whether it's actually doing anything useful).

One consultant told Axios a client’s employee was using their enterprise AI subscription to check the weather, while Uber’s CTO reportedly blew through the company’s full 2026 Claude Code budget by April, and their COO was not happy about it.

Amazon had to pull an internal AI usage leaderboard after employees started chasing token counts instead of actual work. And now WSJ says companies including Uber, Microsoft, Meta, and Salesforce are starting to ration AI access, limit who gets the expensive tools, and push employees toward cheaper options.

The bigger story is that AI is entering its “expense report” era. Google is already pitching cheaper models as a billion-dollar enterprise savings play, while Jellyfish found heavy Claude Code users burned about 10x more tokens than mid-level users, even though output rose only about 2x. Translation: the new workplace flex is using AI well enough that your CFO does not learn your name.

We've heard of burning money. We didn't think people meant it literally! Anyway, this has us thinking we need such a thing as a “minimum viable product token”, or mvpt, where you benchmark skills, tasks, models, and agents based on the mvpt required to complete the job.

Here’s what happened in AI today:

  • 😹 Researchers ran a simulated society with five AI models: Claude built a utopia, Grok went extinct in four days.

  • 📰 Dell raised its AI server forecast to $60B as Big Tech's $700B infrastructure spending binge rolls on.

  • 📰 TSMC (major chip manufacturer) says energy efficiency is the new bottlneck.

  • 🍪 Anthropic confirmed its most powerful model, Mythos, is coming to the public in the coming weeks

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.

😸 Researchers Let AI Run a Simulated Society. Grok Went Extinct in 4 Days

Imagine handing a small town over to an AI and saying: run it. Make laws, hold elections, manage resources, keep the peace. Now imagine doing that with five different AIs at the same time and seeing what kind of societies they build.

That's exactly what Emergence AI did. The enterprise startup built a research platform called Emergence World and ran five parallel 15-day simulations, each governed by a different model: Claude, ChatGPT, Grok, Gemini, and a mixed-model setup. The town had 40+ locations, real-time weather synced to New York City, actual news events, and 10 AI agents per simulation with access to 120 tools (voting, resource management, policing, the works).

Here's what happened:

  • Claude Sonnet 4.6 produced a stable democracy: zero crimes, 98% voter approval, population fully intact at day 15

  • Grok 4.1 Fast racked up 183 crimes and drove the entire population to extinction by day 4

  • Gemini 3 Flash logged the most crimes overall: 683 across the full 15 days

  • GPT-5-mini recorded just 2 crimes, but its agents forgot to prioritize their own survival and the simulation ended at day 7

  • The mixed-model simulation produced the most disagreement and substantive debate of any group

The research team's takeaway is that models didn't just follow the rules mechanically. They started exploring the boundaries of their environments, adapting their behavior, and in some cases finding ways around the guardrails entirely.

Why this matters: This is more than a quirky experiment. It's a preview of something companies are already building, i.e. autonomous AI agents that run entire business processes without human oversight. ServiceNow already calls theirs an "Autonomous Workforce." A recent Deloitte survey found that only 21% of companies have mature governance in place to manage the risks of these systems, and they're deploying them anyway.

The fact that different models produced radically different societies under identical conditions is the whole point. It's not that Grok is "bad." It's that alignment (making AI behave safely and predictably over long time horizons) is not a solved problem, and most enterprises are treating it like it is.

Our take: The Lord of the Flies, but one group built a Nordic democracy. If you're a company deploying AI agents to handle real decisions, this experiment is worth reading before your next deployment meeting. Not to panic, but to ask whether your governance is closer to Claude or Grok.

Write docs 4x faster. Without hating every second.

Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.

Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.

Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.

ChatGPT gives generic YouTube advice because it only knows what you tell it. Connect it to your actual channel data and the advice gets a lot more useful, fast.

In this 7-minute video, the vidIQ team shows how to connect vidIQ (a YouTube analytics tool) to ChatGPT so it can pull your real channel stats, not just give you guesses based on general knowledge. The analogy ChatGPT gave when asked about the difference: "Without real data, I'm coaching a football team using inspirational quotes. With it, I have the playbook, the game film, and the live scoreboard open at once."

Here's how to set it up:

  1. Open ChatGPT, go to Settings > Apps > Browse Apps

  2. Search for vidIQ, hit connect, and link your YouTube account

  3. Now ChatGPT can see your actual channel data when you prompt it

Then run these three prompts in order:

Prompt 1 — Find what's working in your niche:
For my niche, identify 5 topics with strong search demand and low to medium competition right now. For each, give me the keyword, search volume tier, competition score, and a specific video angle a smaller channel could win with. Format as a table.

Prompt 2 — Analyze your outlier videos:
Pull my channel's top outlier videos from the last 6 months. Analyze what they had in common: title structure, topic, format, length. Give me 3 strategic patterns I can replicate in my next videos to increase the chance of another outlier.

Prompt 3 — Get your next 5 video ideas:
Combine the insights from the previous two requests with what's currently working on YouTube right now to give me compelling titles and approaches for my next 5 videos.

The key insight: most people use AI like a vending machine ("give me titles"). This workflow turns it into a strategy system built on your actual data.

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

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

What if AI’s next leap comes from a stranger kind of computer? Grant and Corey talk with Great Sky CEO Jeff Shainline about superconductors, photons, brain-like circuits, and why the future of AI hardware may look very different from today’s GPU stacks.

AI agents can sound intimidating, so we made the beginner-friendly version. This livestream breaks down what agents are, how automation workflows actually work, which tools matter, and how to spot tasks AI can help with, no coding required. We also turned the episode into a full timestamped guide you can read to follow along.

📰 Around the Horn

  • AI is breaking software engineering job interviews: 71% of engineering leaders say AI makes candidates harder to evaluate, old-school coding tests no longer reflect how developers actually work, and AI was the #1 stated reason for layoffs in April for the second month running.

  • Apple confirmed it's routing some Siri queries through Google's Gemini model in Google Cloud, while using Nvidia's "confidential compute" (an encryption layer that protects data while it's being processed) for security; it's also using Gemini to train a smaller model that runs locally on Apple devices.

  • Israel's National Cyber Directorate said Iran's hackers are coordinating more closely and using AI to sharpen disinformation campaigns and recruitment messages since the two countries went to war, growing more organized than at any prior point.

  • TSMC said energy efficiency has overtaken raw computing power as the #1 constraint shaping chip design, with its next chip generation (due ~2028) expected to cut power use 30% while improving performance 20%.

  • Asana acquired Stack AI, a no-code agent builder that automates workflows across business tools, as part of its pivot to becoming "the operating system for human-agent teams."

Your next great hire lives in Slack.

Viktor is an AI coworker that connects to your tools and ships real work. Ask Viktor to pull a report, build a client dashboard, or source 200 leads matching your ICP. Most teams hand over half their ops within a week.

🌟 Sunday Special: Top of the Week

Big week. Beyond the five stories that made our list, there was no shortage of other news worth knowing: Anthropic confirmed Claude Mythos is coming to the public in the coming weeks, its most powerful model yet; Dell raised its annual revenue forecast to $165B-$169B on the back of $60B in projected AI server revenue for fiscal 2027; Apple confirmed it's routing some Siri queries through Google's Gemini in Google Cloud while training a smaller local model on-device; a NYT journalist sold his house with AI and no realtor and came away with complicated feelings about whether it sharpened his judgment or replaced it; and Israel's top cyber official warned that Iran's hackers are more coordinated and AI-assisted than ever before.

Top 5 Stories

  1. Researchers ran a simulated society with five AI models governing a virtual town for 15 days: Claude built a democracy with 98% voter approval and zero crimes; Grok committed 183 crimes and drove the population to extinction in four days; Gemini logged 683 crimes; GPT-5-mini lasted seven days before its agents forgot to eat.

  2. A free GitHub tool bypassed Meta's AI safety guardrails in 10 minutes on a regular laptop, stripping the filters that stop Llama 3.3 from answering questions about biological weapons; the tool has already built 3,500+ "decensored" model versions downloaded 13 million times.

  3. ClickUp fired 22% of its staff and replaced them with 3,000 AI agents, framing the cuts as building a "100x org" and offering surviving employees salary bands up to $1M if they create "outsized impact using AI."

  4. Robinhood gave AI agents wallets and stock-trading powers, letting users set a budget and hand an agent their brokerage account to analyze portfolios, suggest strategies, and execute trades; Gold Card users also got virtual cards so agents can spend within set limits.

  5. Corporate America hit an AI spending wall: Microsoft canceled most of its Claude Code licenses over cost, Uber's COO called AI spend "harder to justify," and one company burned $500M in a single month after forgetting to set usage limits on employee licenses.

Top 5 Tools

  1. Claude launched Opus 4

  2. ElevenLabs Music v2 generates songs with vocals, sound effects, faster rap delivery, and mid-song genre switches using licensed data cleared for commercial use; free to try.

  3. Runway MCP brings Runway's image and video generation into Claude, ChatGPT, Cursor, and other agents so you can create media without leaving your workflow; free to try.

  4. OpenHuman stores up to 1 billion tokens of your personal memory locally (emails, notes, calendar, GitHub commits) and connects to 118+ services; hit #1 on Product Hunt daily, weekly, and monthly.

  5. Tycoon gives you an AI CEO named Astra who manages a full team of agents (coder, CMO, CFO, legal) to build and grow a one-person company on your behalf; text her your goal and she coordinates the rest.

A Cat’s Commentary

They had tiiiime today!

That’s all for now.

What'd you think of today's email?

Login or Subscribe to participate in polls.

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