Something crystallized this week. The same technology making coders obsolete is now being conscripted—and one of the most consequential governance fights in AI history is playing out in real time. Meanwhile, the guy who built Meta's AI research empire just raised a billion dollars to prove the whole approach is wrong. Let's get into it.

The Anthropic-Pentagon Fight Is About Something Much Bigger

Dwarkesh Patel published a long, serious essay this week on the Department of War's move to blacklist Anthropic as a supply chain risk. And his argument cuts much deeper than the headline conflict.

The backstory: Anthropic refused to remove usage restrictions on their models—specifically around autonomous weapons and mass surveillance. The Pentagon responded by threatening to designate them a national security risk, which would effectively force Amazon, Google, and Nvidia to cut Anthropic out of any defense-adjacent work.

Dwarkesh's core argument is that the government made a tactical error. If you're worried about private companies having veto power over military AI, the answer is to not do business with them—not to destroy them for refusing to comply. That second move plants a very dangerous precedent.

He runs the math on mass surveillance costs. There are one hundred million CCTV cameras in America. Processing a frame every ten seconds, at current multimodal model pricing, runs about thirty billion dollars annually today. With AI costs dropping roughly ten times per year, by 2030 that number approaches a few hundred million. The technical bottleneck disappears. What remains is political will—and norms.

That's why the amicus briefs matter. Thirty tech employees, including Google's chief scientist Jeff Dean, lined up behind Anthropic. Microsoft filed formally. These aren't just expressions of solidarity. They're signals about where industry norms are forming.

The timing of Anthropic's other move is worth noting. They launched the Anthropic Institute this week—a thirty-person group merging their red team, societal impacts, and economics research under Jack Clark. The mandate: study and communicate AI's disruption in real time. If what some researchers believe is true—that seventy to ninety percent of code for future Claude models is already written by Claude itself—then having an institute already studying that feedback loop looks prescient rather than reactive.

Dwarkesh ends somewhere uncomfortable. He argues that even courageous corporate refusals won't hold the line. Open-source models will eventually be capable enough for mass surveillance. The only durable answer is political: establishing the same norms around AI-enabled surveillance that we established around chemical weapons after World War One. Not corporate policy. Law.

Agents Are Eating Knowledge Work

Andrej Karpathy made an interesting move this week. Pushed back on the "age of the IDE is over" framing. His counter: "we're going to need a bigger IDE." The unit of work isn't a file anymore. It's an agent.

That framing captures what's actually happening. Coding agents—the tools that were just generating functions eighteen months ago—are rapidly expanding their scope upward into general knowledge work.

Replit is the clearest example. Agent 4 launched this week, and the product is genuinely unrecognizable from the "coding with AI tacked on" platform of two years ago. Canvas, apps, sites, slides, video—a full productivity suite. The underlying logic: now that coding is approximately solved, the same agent infrastructure moves up the stack.

Perplexity made a similar move with Personal Computer—an always-on local agent that runs on a Mac mini, with persistent access to your files, apps, and sessions, controllable remotely. They're explicitly framing it as the safer alternative to OpenClaw—with a kill switch, tracked activity, and sign-off requirements for sensitive tasks.

The pattern Swyx at Latent Space identified is worth internalizing. Across 2026 so far, you see coding agent builders—Pi, Claude Code, Replit—all expanding scope toward general knowledge work. The infrastructure built for code generation turns out to generalize. Writing, research, data analysis, presentation—all the same underlying capabilities.

What's changing in the engineering discussion is a shift from "which model?" to "how do you build the harness?" LangChain added autonomous context compression to their deep agent framework this week—so models can compact their working memory at task boundaries rather than hitting hard token limits. That's the kind of infrastructure work that makes the difference between agents that get stuck and agents that actually complete things.

The Open Model Efficiency Race

NVIDIA dropped Nemotron 3 Super this week—a one hundred twenty billion parameter model with only twelve billion parameters active at any time. That's a mixture-of-experts—or MoE—architecture, where the model routes each token through a small subset of specialized sub-networks.

The interesting technical story isn't the raw capability numbers. It's the inference speed. NVIDIA is claiming up to two-point-two times faster than competing open models in FP4 precision. Two mechanisms explain it.

First, multi-token prediction—the model makes provisional guesses about several future tokens simultaneously, then verifies them, exploiting otherwise-idle GPU compute at small batch sizes. Second, a dramatically smaller key-value cache—the memory structure that stores context during generation. Nemotron uses roughly eight thousand bytes per token in BF16 precision. Qwen 3.5-122B uses about twenty-four thousand. For long-context workloads, that three-to-one difference compounds.

Community benchmarks on the new M5 Max confirmed the pattern. Local inference practitioners are getting up to around five hundred tokens per second on the model with the right setup. Day-one support landed across the major inference stacks—vLLM, llama.cpp, Ollama—so this one reaches practitioners fast.

Yann LeCun's Billion-Dollar Bet

Here's the standalone story that doesn't fit neatly elsewhere but deserves your attention.

Yann LeCun—Turing Award winner, former chief AI scientist at Meta, arguably the most credentialed critic of large language models—left Meta four months ago and just raised a billion dollars for a new company called AMI. Advanced Machine Intelligence Labs.

His argument hasn't changed. LLMs learn to predict the next word. They're impressive. They're not intelligent. They don't understand physical reality. They hallucinate because they're generating plausible surfaces rather than modeling underlying structure.

His alternative is JEPA—the Joint Embedding Predictive Architecture—a framework he first proposed in 2022. The idea is to learn abstract representations of how the world works, not pixel-perfect or word-by-word predictions. Systems that understand reality the way animals do: through embodied experience, not text.

AMI's founding team is drawn almost entirely from Meta AI research. His stated timeline: corporate partnerships within one to two years, "fairly universal intelligent systems" within three to five. The geographic bet is also deliberate—Paris headquarters, explicitly positioned as a European counterweight to American and Chinese AI giants.

Here's why this matters regardless of whether you believe LeCun is right. The investors who put a billion dollars behind him—Bezos Expeditions, Nvidia, Toyota, Samsung—are hedging. They're not betting LLMs fail. They're betting that whatever comes next might not look like GPT. And if AMI does produce something that works at scale with genuine physical-world understanding, it reshapes a lot of assumptions about where the current trajectory ends.

The Thread Running Through Everything

Pull back and the week's stories connect. The Anthropic-Pentagon fight is fundamentally about who gets to define what AI systems should and shouldn't do—at the moment when those systems are becoming genuinely indispensable infrastructure. Karpathy's "bigger IDE" thesis is a description of that infrastructure expanding. The open model efficiency race means that infrastructure becomes available to everyone, including actors you can't negotiate redlines with.

LeCun is betting all of this leads somewhere wrong. The Anthropic Institute is betting we need to study where it leads, urgently, before we arrive.

Everyone is running. Not everyone agrees on the destination.