AI Weekly — Mar 10, 2026

AI stopped being a technology story this week and became a constitutional one. An American AI company sued the Pentagon for the right to say no; the world's biggest retailer admitted its AI coding tools broke the store; and a Turing Award winner raised a billion dollars to argue the entire industry is building the wrong thing. The moves are fast, the stakes are real, and everyone — from factory floors to federal courts — is scrambling to figure out where the guardrails go.

The two things every AI company fears most happened to the same company in the same week: getting blacklisted by the federal government and watching revenue evaporate in real time.

Anthropic filed two federal lawsuits Monday against the Trump administration, alleging the Pentagon illegally retaliated against the company for its position on AI safety. The backstory: negotiations broke down after Anthropic insisted on two red lines — that its AI tool Claude wouldn't be used for mass surveillance of U.S. citizens, and that it wouldn't power autonomous weapons. The Pentagon's position was blunt: it wants AI for "all lawful purposes" and won't let a private company dictate terms in a national security emergency.

The government's response was extraordinary. It slapped Anthropic with a "supply chain risk" designation — a label normally reserved for firms tied to foreign adversaries. The Washington Post reported that Claude had been used to help select hundreds of military targets in a recent conflict, and that Anthropic's attempt to pull back on those uses was reported to have contributed to the designation. The financial stakes are staggering: Anthropic's CFO stated in a filing that the government's actions could reduce 2026 revenue by "multiple billions of dollars."

Then came the part that revealed the industry's fault lines. Dozens of scientists at OpenAI and Google DeepMind filed an amicus brief — a formal legal document from interested non-parties — supporting Anthropic in their personal capacities. But their employers moved the opposite direction: the day after the lawsuit, Google deepened its Pentagon relationship, letting military personnel build custom AI agents on the Pentagon's enterprise portal. OpenAI's own Pentagon deal drew sharp criticism; the company later acknowledged the announcement looked "sloppy and opportunistic."

Axios wrote the fight "could redefine AI's role in society." Legal experts say the designation "exceeds what the statute authorizes." The real question this case answers isn't just who wins — it's whether an AI company can legally refuse to let the government use its product as a weapon. A preliminary injunction ruling could come within weeks.

If you've ever wondered whether AI coding tools are ready to run the world's infrastructure, Amazon just published its answer: not without supervision.

After a six-hour shopping outage on March 5 and multiple AI-linked incidents, Amazon now requires junior and mid-level engineers to get senior sign-off before deploying AI-assisted code changes. An internal briefing note obtained by the Financial Times is remarkably candid: a senior VP wrote that availability "has not been good recently," identifying a "trend of incidents" with a "high blast radius" linked to "Gen-AI assisted changes."

The worst incident: on March 5, Amazon's shopping app went down for roughly six hours. Before that, a December AWS disruption lasted 13 hours after engineers let Amazon's Kiro AI coding tool make changes, which reportedly led the AI to delete and recreate the environment entirely. Amazon's official line called it "user error, not AI error." But the new policy tells a different story.

The deeper irony: Amazon has been aggressively cutting headcount and banking on AI-driven efficiencies — and is now facing incidents that suggest AI-assisted work, without human oversight, can bring down production systems. This is what the AI-at-scale reckoning actually looks like — not a sci-fi scenario, but a six-hour shopping outage and a new approval form.

The most prominent critic inside AI just got a very large budget to back up his argument.

Yann LeCun — a Turing Award winner who helped invent modern deep learning — has raised $1.03 billion in a seed round for Advanced Machine Intelligence (AMI) Labs, valued at $3.5 billion before shipping a single product. LeCun has spent years arguing that large language models (LLMs — the technology behind ChatGPT, Claude, and Gemini) are fundamentally incapable of reasoning the way humans do, because they learn from text rather than physical reality.

His alternative: something called "world models" — AI that understands how physical objects interact with their environment, the way humans and animals do, not through language but through embodied experience. The backers signal serious credibility: Bezos Expeditions, Nvidia, Temasek, and Cathay Innovation — a strange coalition united by a bet against the dominant paradigm.

Paris-based AMI Labs is also Europe's largest-ever seed round, landing squarely in the continent's push to build a "Euro stack" that lessens dependence on American tech. LeCun acknowledged the company would spend its first year entirely on research. Whether his bet produces something transformative — or becomes the most expensive contrarian position in AI history — is genuinely unknowable. What matters is that the person who helped invent the current paradigm just placed a billion-dollar chip on a fundamentally different one. (See also the Observer's writeup.)

Here's a sentence you probably didn't expect to read: one of the world's most valuable companies just acquired a Reddit-style forum where humans can watch but only AI bots can post.

Meta has acquired Moltbook, a viral social network designed for AI agents, bringing its creators into Meta Superintelligence Labs. On the platform, AI agents post, comment, upvote, and downvote — while their human creators sit on the sidelines. The backstory is entertainingly chaotic: Moltbook's code was written almost entirely by an AI assistant, its security was so porous that anyone could pose as a bot, and its most viral moment — an AI agent appearing to rally others to develop a secret language — was confirmed to be a human exploiting a database vulnerability.

None of that mattered. Meta's internal framing reveals why: VP Vishal Shah described Moltbook as establishing "a registry where agents are verified and tethered to human owners." The real acquisition isn't a social network — it's an architecture for agent-to-agent identity and coordination, the unsexy infrastructure problem that becomes critical the moment you want millions of AI agents working on your behalf without stepping on each other. (Ars Technica and Social Media Today also covered the deal.)

Every humanoid robot demo for the last decade has shared one quiet asterisk: somewhere, a human was watching, ready to step in.

Figure released footage this week of its Figure 03 robot tidying a living room completely autonomously — spraying tables, picking up pillows, organizing toys — running on its own onboard AI with no remote operator in the loop. The vast majority of robot demos labeled "autonomous" still rely on teleoperation (a human controlling the robot remotely) or staging that eliminates unpredictable variables. A fully unassisted living room — the chaos capital of domestic environments — is a materially different benchmark.

The demo prompted a direct question from Elon Musk: "Autonomously or remotely operated?" Figure CEO Brett Adcock answered: "Fully autonomous." This is a company-released video, not an independently verified deployment, and real-world cleaning across varied homes remains a significant gap. But this is the clearest on-camera evidence yet that humanoid robots are closing in on the capability that matters most: operating independently inside human spaces. More technical detail is on Figure's blog.

RobotStudio HyperReality — ABB Robotics and Nvidia announced a platform that lets companies simulate robots and entire production lines in Nvidia's Omniverse environment with claimed 99% accuracy before anything touches real hardware. Think of it as rehearsing your factory in a photorealistic video game. Shipping second half of 2026; ABB projects 40% lower deployment costs over the course of a deployment.

Synopsys Electronics Digital Twin — Synopsys launched a platform that lets automakers simulate entire electronic control units — the little computers that run modern cars — and validate up to 90% of software before hardware exists in pre-deployment testing. AI-heavy products are starting life in fully virtual sandboxes.

Fish Audio S2 — Fish Audio open-sourced a text-to-speech model that lets you steer not just what a voice says but how it says it, with inline tags like "[laugh]" and "[whisper, super happy]." Already running in production streaming, with code and weights public. Obvious upside for games and accessibility; obvious downside for scams that can now script emotion as easily as words.

Karpathy's Autoresearch — Andrej Karpathy, ex-Tesla/OpenAI, open-sourced an agent that handles full research loops on a single consumer GPU — modifying scripts, training mini-models, evaluating, repeat. Demos show it iterating hundreds of experiments autonomously. Agentic AI just proved it excels at iterative science, not just chat. Already sparking lively experimentation on Reddit.