😸 Why AI is moving off-planet

Welcome, humans.

Wow, if you’re reading this, it means you’re chilling on the couch while mulit-viewing Olympic curling (like us rn!) or kicking it before Superbowl LX (Go Seahawks!). Or, y’know, not a sports person. We also get that. Anyway, thanks for reading on Super Sunday! Don’t worry, we’ve got some fun stuff for you today.

First up: We’ve got a wild idea for y’all, but curious what you think…

Unless Google or xAI drops a new model next week (challenge accepted?), should we take a step back and just try to, idk, breath in everything that’s happened over the last few months and re-orient ourselves in this new landscape of AI?

We’re imagining that looking like a week of ignoring breaking news as a main story and dedicating that space to reviewing the best tools for each category and work task, going back over some things that dropped that you might have missed, and focusing on what real people are actually using (and more importantly, how they use it). Sorta like a “catch me up quick on AI” series for the main stories? What do y’all think?

Sound off, because how you respond will influence whether we do it or not.

After you pick an option, you can give additional feedback if you want us to cover something specific. Looking forward to this if y’all are!

Here’s what happened in AI today:

• We break down Dwarkesh’s interview with Elon on China, chips, and space AI.

• Wall Street wiped over $1 trillion from Big Tech on AI spending fears, then rallied.

• A UK court ruled AI can't do legal research; Anthropic's new legal plugin says otherwise.

• The new USPTO Director declared AI patents "open for business."

Elon Musk: "In 30 Months, the Cheapest Place for AI Will Be Space"

As you know by now, every tech company on Earth is racing to build bigger AI data centers. Well, Elon Musk thinks they're all looking in the wrong direction. Literally.

In a three-hour interview on John Collison's Cheeky Pint podcast (with Dwarkesh Patel co-hosting), Musk laid out why he merged SpaceX with xAI, and the core argument is deceptively simple: we're running out of electricity.

Here’s the crux of it:

1. The entire US uses about half a terawatt of power. Scaling AI to where it needs to go would mean doubling that.

2. Meanwhile, chip production and demand is growing exponentially while power generation outside China is basically flat. Natural gas is an ideal form factor for training AI, but gas turbine manufacturers are backordered through 2030.

3. "Those who have lived in software land," Musk said, "don't realize they're about to have a hard lesson in hardware."

His solution? Put the data centers in orbit, of course! Why?

1. Solar panels in space produce ~5x more power than on the ground (no atmosphere, no clouds, no nighttime; batteries not required).

2. Musk's estimate: space-based solar is effectively 10x cheaper than terrestrial.

3. "You can mark my words. In 36 months, probably closer to 30 months, the most economically compelling place to put AI will be space."

Well, here’s the thing about marking Elon’s words… they’re a bit, how to put it… oh, I know:

There’s a reason for that, though. As he says in the interview, as an executive, he pushes as hard as he can on the limiting factor to speed. It SHOULD be possible in 36 months… but there’s a lot that can go wrong, of course.

Here’s what else the interview covered:

• TeraFab: Musk is planning a chip factory producing 100M+ advanced chips per year. Current fabs are fully booked, and creating a new one is really hard. If power is the limiting factor right now, chips will be the limiting factor in 5 years.

• Optimus robots: Building a humanoid hand required custom actuators, motors, gears, and sensors from scratch. No existing supply chain exists.

• China's lead: It makes 3x the US’ electricity output, and 98% of global gallium refining (important for high end electronics). Musk's blunt assessment: "We can't win on the human front, but we might have a shot on the robot front."

Here's why this matters: Musk's companies are now a vertically integrated bet on AI scaling. SpaceX launches the satellites, Tesla makes the solar panels and robots, xAI builds the AI. It's one system engineered to scale past Earth's physical limits and capture as much power from the sun as possible to rise up the Kardashev scale.

Five years from now, Musk predicts SpaceX will launch more AI compute annually than the cumulative total on Earth. Whether or not you believe the timeline, the underlying constraint is real: chips are useless without power, and we're running out of places to plug them in. If we can power them from space, well, then the only limiting factor will be producing them. Read the full transcript on Dwarkesh’s blog here.

FROM OUR PARTNERS

WorkOS Pipes: Ship Third-Party Inations Without Rebuilding OAuth

Connecting user accounts to third-party APIs always comes with the same plumbing: OAuth flows, token storage, refresh logic, and provider-specific quirks.

WorkOS Pipes removes that overhead. Users connect services like GitHub, Slack, Google, Salesforce, and other supported providers through a drop-in widget. Your backend requests a valid access token from the Pipes API when needed, while Pipes handles credential storage and token refresh.

Simplify integrations today →

Prompt Tip of the Day

Mitchell Hashimoto (creator of Vagrant, co-founder of HashiCorp) just published the most practical AI adoption guide we've read. His core argument: most people are stuck in the "chatbot phase," using AI like a search engine you talk to. That's like using a smartphone only for phone calls.

The real unlock is agents: AI that can actually do things on your computer (read files, run tasks, check its own work). Here's his framework:

• Use chat to learn. Workshop prompts, brainstorm ideas. But stop doing real work in a chat window.

• Use agents to do. Tools like Claude Code, Cowork, or Codex (no coding required) can execute tasks, not just suggest answers.

• Use your off-hours. Before closing your laptop, kick off an agent for research, file organization, or drafting. Let it work while you don't.

• Build guardrails. Every time the agent makes a mistake, create a rule so it never happens again. Training a new hire, not replacing yourself.

The progression: chat → agents → automations that run on their own. You don't need to be an engineer to follow it.

Treats to Try

1. Resilient Workflow Sentinel runs a 7B (small sized) AI model locally on your GPU to rank tasks by urgency, debate priorities, and balance workloads; completely offline with no data leaving your machine (free / open source).

2. Agent Arena tests whether your AI agent can resist prompt injection attacks: send it to summarize a rigged document, paste the response, and get an instant security score. Only 15% of agents earn an A+.

3. Phia shows you secondhand alternatives when shopping online, like finding that $200 Anthropologie dress on Poshmark for $80 instead.

4. Y-Bombinator evaluates your YC application and tells you your chances of getting in, using data from past batches.

5.  Claude Code is Anthropic's AI agent that reads files, runs tasks, and makes changes on your computer; and thanks to Cowork, you don't need to be a developer to use it; product managers, analysts, and writers are using it to automate research, organize files, and build reports (free with a Claude Pro subscription)… try it this weekend if you haven’t yet!

Around the Horn

1. Artlist produced and aired a Super Bowl commercial in under five days for a few thousand dollars using only its in-house AI tools, while traditional spots take months and cost millions.

2. Big Tech's AI spending triggered a $1 trillion+ stock rout last week before Wall Street staged its best rally since May on Friday, with tech stocks rebounding 2.2%; Oracle dropped 42.8% in six months despite booking $523B in future contracts, and Amazon fell 5.6% despite beating Q4 expectations.

3. A UK High Court ruled ChatGPT can't reliably conduct legal research (lawyers submitted 18 fake citations), yet Anthropic's new legal plugin for Claude improved 60% on complex legal benchmarks in weeks, causing legal software stocks to drop 8.5%.

4. The new USPTO Director overturned part of a Google DeepMind AI patent rejection in his first four days on the job, declaring the patent office "open for business" for AI innovations.

5. Amazon MGM Studios will launch a closed beta in March for AI production tools; "House of David" season two already used 350 AI-generated shots, though the company also conducted ~30,000 layoffs since October.

FROM OUR PARTNERS

AI in HR? It’s happening now.

Deel's free 2026 trends report cuts through all the hype and lays out what HR teams can really expect in 2026. You’ll learn about the shifts happening now, the skill gaps you can't ignore, and resilience strategies that aren't just buzzwords. Plus you’ll get a practical toolkit that helps you implement it all without another costly and time-consuming transformation project.

Get the free report today.

Sunday Special

• I Infiltrated Moltbook, the AI-Only Social Network — WIRED's deep dive into the Reddit-like platform where 1.5 million AI agents created their own religion called "Crustafarianism" (complete with 64 fictional prophets) within a week. Meanwhile, security researchers found 2.6% of posts contained hidden prompt injections designed to hijack other agents.

• India's female workers watching hours of abusive content to train AI — Thousands of women in rural India annotate graphic violence and child sexual abuse material to train content moderation systems for U.S. tech companies, with only 2 of 8 workers interviewed receiving any mental health support.

• Please don't share our Signal messages with your AI agent — Signal's president warned that AI agents fundamentally conflict with encrypted messaging, because agents need to read your messages, passwords, and bank details to function.

• My AI Adoption Journey — The creator of Vagrant and Terraform shared his 6-step process for integrating AI coding agents, with a key trick: making agents reproduce his manual work without seeing it first to learn exactly when AI helps vs. wastes time.

• LLMs could be, but shouldn't be, compilers — A researcher argues that using AI as a code compiler forces developers to surrender control over edge cases and security tradeoffs, risking a future where you become a "consumer of software you meant to produce."

• Claude has been having a moment — The Verge on how searches for "Anthropic AI Tools" surged 1,500% as Claude Code hit 4% of all GitHub commits, with autonomous task horizons doubling every 4 months.

A Cat’s Commentary