Today AI isn't just a tech story. It's a power story.

The U.S. government just tested whether an AI company can actually say no to a direct threat. And from another corner, a small lab quietly dropped what many are calling the GPT-3 moment for computer use.

Two stories that seem unrelated, but both point to the same thing: AI is entering a phase where the stakes are no longer theoretical.

Defense Secretary Pete Hegseth met directly with Anthropic CEO Dario Amodei and delivered three options: remove Claude's safety guardrails for military use, accept being labeled a "supply chain risk" (a designation normally reserved for foreign adversaries), or be forced to comply under the Defense Production Act.

Anthropic refused all three.

This is not a dispute about features. Anthropic has long held firm on two specific limits: Claude will not be used in fully autonomous weapons without a human in the loop, and it will not be deployed for mass surveillance of American citizens. Meanwhile, xAI signed a deal with the military after agreeing to "all lawful purposes" use cases. OpenAI and Google are both being fast-tracked for classified network access.

This is the first time a government has openly used legal threats to pressure an AI lab into dropping its guardrails. If the Defense Production Act gets invoked successfully here, the precedent it sets goes far beyond Anthropic. It signals to every lab: push back on the government and you risk losing everything.

The real fight is not about Claude. It is about who gets to draw the red lines on AI.

Standard Intelligence just released FDM-1, the first model trained on 11 million hours of screen recordings to learn how to operate a computer the way a human does: continuous mouse movements, keystrokes, 3D modeling in Blender, and even driving a real car after less than one hour of fine-tuning. The technical breakthrough is their video encoder, which compresses nearly 2 hours of 30 FPS video into just 1 million tokens, making it 100x more efficient than OpenAI's encoder.

Before FDM-1, all computer use agents were limited to a few seconds of context and required expensive human-labeled data. FDM-1 breaks both constraints at once by learning unsupervised from internet video and handling tasks that span hours. It is the first computer use model pretrained at internet scale, a milestone comparable to what GPT-3 was for language in 2020.

If LLMs turned text into fuel for AI thinking, FDM-1 is doing the same for computer actions. When a model can watch you work and learn to replicate it, the line between "AI that assists" and "AI that replaces" gets thin fast.

Not today. But the direction is clear.

TechCrunch reports a new pattern: AI-native startups are reaching multimillion-dollar revenue milestones within months of launching, faster than any previous generation of startups. AI is collapsing the distance between idea and real revenue at an unprecedented pace.

→ You no longer need a 50-person team and a Series B to prove product-market fit.

Anthropic unveiled a new enterprise agents program that lets companies deploy pre-built agents for financial research, engineering specs, and design tasks directly inside their existing workflows. It is Anthropic's clearest move yet into the SaaS enterprise market.

→ If agents actually deliver, an entire category of B2B software is in the crosshairs.

Despite all the hype, OpenAI's own COO acknowledged that AI has not meaningfully embedded itself into the core operations of most large enterprises. The gap between "trying AI" and "running AI in production" remains wide.

→ An honest admission, and exactly the opening Anthropic and others are racing to close.

*Falconer: Self-maintaining docs for engineering teams. Instead of manually writing and updating documentation, Falconer syncs with your codebase and keeps docs accurate automatically. Built for anyone tired of outdated technical docs causing real problems.

*RunAnywhere: Run AI models locally on any device, no cloud, no internet required. A solid option for anyone handling sensitive data or wanting to stay off external servers entirely.

💡 How to use this today: If you work with an engineering team, try demoing Falconer on one small project. Outdated docs tend to be ignored until they cause a real bug. And if you occasionally run internal files or client data through AI tools, RunAnywhere is worth a look purely from a privacy standpoint.

Anthropic just turned down $200M and the risk of a government blacklist to hold a principle. On the same day, a small lab showed that AI can learn to do anything simply by watching. The question is not where the technology is going. The question is: when AI can do everything, what will humans choose to keep for themselves?

👉 Forward it to someone trying to keep up with AI.

Read online: leftcurve.beehiiv.com