AI in 15 — March 27, 2026

The Pentagon just got told by a federal judge that branding an American company a potential saboteur for disagreeing with the government is, and I quote, "Orwellian." Forty-three pages of judicial fury. A hundred-dollar bond. That's the court's way of saying your case is worthless.

Welcome to AI in 15 for Friday, March 27, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Friday, Marcus. And what a week to close out. Judge Rita Lin has officially ruled in Anthropic's favor with a preliminary injunction blocking the Pentagon's blacklisting. A startup called Symbolica just embarrassed every frontier model on the new ARC-AGI-3 benchmark using agents instead of raw reasoning. Intel dropped a 32-gigabyte GPU for under a thousand dollars aimed squarely at local AI. Claude Code now runs tasks in the cloud while you sleep. And a school used AI to flag 1984 for removal from its library. You cannot make this up. Let's get into it.

Judge Lin blocks the Pentagon's Anthropic blacklist in a blistering 43-page ruling.

Symbolica scores 36 percent on ARC-AGI-3 while the best LLMs can barely manage one percent.

And Intel launches a 32-gig GPU for local AI inference at 949 dollars.

Marcus, we've been covering the Anthropic-Pentagon case all week. Monday the hearing, Thursday the judge's strong language, and now Friday we have an actual ruling. Walk us through what happened.

Judge Rita Lin issued a preliminary injunction. Forty-three pages. And Kate, the language is extraordinary for a federal court opinion. She wrote that nothing in the governing statute supports the "Orwellian notion" that an American company can be branded a potential adversary for expressing disagreement with the government. She also pointed out that if the Pentagon's concern was operational, they could simply stop using Claude. Instead, she said, these measures appear designed to punish Anthropic.

And the bond was set at a hundred dollars. That's basically pocket change for a legal proceeding this significant.

That's the judicial equivalent of laughing the government out of court. A nominal bond signals the judge thinks the government has almost no chance of prevailing on the merits. The injunction blocks the supply chain risk designation, which was the nuclear option the Pentagon used. That designation is normally reserved for companies like Huawei and Kaspersky. Foreign adversaries. Applying it to an American company because they wouldn't remove ethical guardrails on autonomous weapons and mass surveillance was unprecedented.

There's an NPR investigation that dropped this week that adds crucial context here. Federal agencies are already buying Americans' data in bulk from data brokers, bypassing warrants entirely.

And that's exactly what makes Anthropic's stance more than corporate ethics theater. Dario Amodei warned that commercial data combined with AI creates the ability to assemble a comprehensive picture of any person's life automatically and at massive scale. The NPR piece shows that ICE, the FBI, and the Department of Defense are already buying location data, browsing records, personal information. They have the raw ingredients. What they need is the AI processing power to make it actionable at scale. That's what Anthropic refused to provide.

So when Anthropic drew those red lines, they weren't being hypothetical.

Not remotely. The surveillance infrastructure exists. The data pipelines exist. The missing piece was a powerful AI system willing to process it all without restrictions. And when Anthropic said no, the government tried to destroy them for it. Judge Lin just said that's not how this works. The implementation is delayed one week to allow an appeal, and this is almost certainly heading to the Supreme Court eventually.

Meanwhile, the QuitGPT movement has over two and a half million supporters, Claude is still riding high on the App Store, and OpenAI is the one with the Pentagon contract Anthropic refused.

The market has spoken pretty clearly on which side of this it prefers. But the legal precedent matters more than the App Store rankings. This ruling establishes that the government cannot weaponize supply chain security designations against domestic companies for ethical disagreements. Every AI company negotiating government contracts is watching this case.

From the courtroom to the benchmark lab. We covered ARC-AGI-3 yesterday when it launched and AI scored about twelve percent. But Marcus, something remarkable happened within twenty-four hours. A startup called Symbolica scored thirty-six percent using a completely different approach.

This is a genuinely striking result. To recap, ARC-AGI-3 tests interactive abstract reasoning. No instructions, no descriptions. You have to figure out what to do through trial and error. When we reported yesterday, the best frontier models were scoring around twelve percent in the standard evaluation. But Symbolica used their Agentica SDK, an agentic framework, and hit thirty-six percent on the public problem set within twenty-four hours. Claude Opus 4.6 using chain-of-thought reasoning scored 0.2 percent on the same problems. GPT-5.4 managed 0.3 percent.

Wait. Zero point two versus thirty-six? That's not an incremental improvement. That's a completely different universe.

A hundred and eighty times better performance. And here's the kicker. It cost about a thousand dollars in compute. The Opus baseline that scored a hundred and eighty times worse cost nearly nine thousand dollars. So it's not just dramatically better, it's dramatically cheaper. The difference is the agentic architecture. Instead of trying to reason through the puzzle in one shot, the Agentica framework runs autonomous loops that explore, hypothesize, test, and refine solutions iteratively.

There are caveats though, right?

Important ones. This was on the public development set of twenty-five problems, not the hundred and ten private evaluation problems. The ARC-AGI-3 paper explicitly notes the public set is materially easier. And the result doesn't qualify for the official leaderboard because it uses an external harness. But even with those caveats, a hundred and eighty X performance gap between agentic and chain-of-thought approaches on the same task is a data point you can't ignore.

And humans still score a hundred percent.

Which keeps AGI claims firmly in check. But the takeaway for the industry is clear. The frontier of AI capability is moving from bigger models to smarter scaffolding around existing models. That's where the investment thesis is heading.

Intel just launched something that I think a lot of AI developers have been waiting for. The Arc Pro B70. Thirty-two gigabytes of VRAM for 949 dollars. Marcus, why does that number matter?

Because VRAM is the bottleneck for local AI inference. Nvidia's consumer cards top out at twenty-four gigs with the RTX 4090, which costs over two thousand dollars used at this point. Their professional AI cards with more memory cost five to ten times that. Intel just put thirty-two gigs of GDDR6 on a workstation card for under a thousand dollars with six hundred gigabytes per second of memory bandwidth.

And thirty-two gigs means you can run models that simply won't fit on consumer hardware.

Exactly. A quantized seventy-billion parameter model needs roughly thirty-two to forty gigs of VRAM. On a twenty-four gig card, you're stuck with smaller models or aggressive quantization that hurts quality. On the B70, you can run the full model. For developers and researchers doing local inference, this is a compelling value proposition.

Intel's GPU track record has been rocky though.

That's the legitimate concern. Previous Arc launches had availability issues and driver problems. This is a workstation card, so driver stability matters even more than on consumer hardware. But if Intel delivers on availability and software support, they've carved out a real niche. Not competing with Nvidia on gaming or training. Competing on the specific use case of local inference where memory capacity matters most and price matters second.

Anthropic shipped a new Claude Code feature this week. Cloud-based scheduled tasks. You set a cron schedule, point it at a repo, write a prompt, and Claude wakes up on Anthropic's servers to do the work. Marcus, this feels like a step toward AI agents that just run in the background.

That's exactly what it is. The existing loop command runs tasks locally, which means your machine has to be on. Cloud scheduled tasks run on Anthropic's infrastructure twenty-four seven. You could have Claude review your codebase every morning at six AM, check for dependency vulnerabilities weekly, or run analysis tasks overnight. The current limits are conservative. Three daily sessions even on the highest tier plan. And there are restrictions on outbound HTTP and no screenshot capability.

But the direction is clear.

Unambiguous. We're moving from AI as an interactive tool you invoke toward AI as a background process that works autonomously on your behalf. The restrictions tell you Anthropic is being careful about compute costs and potential misuse. But this is the thin end of a very significant wedge.

Quick hit on the Reco story. A security company used AI to rewrite a JavaScript library in Go in about seven hours, spending four hundred dollars in tokens, and claims it'll save them roughly three hundred thousand a year in compute costs. Marcus, is this as impressive as it sounds?

The result is real. They eliminated a fleet of Node.js Kubernetes pods that their Go pipeline was calling via RPC to evaluate JSON expressions. Pure Go implementation, twenty-five to ninety X speedups on complex operations. But the Hacker News crowd correctly pointed out that the expensive part wasn't the rewrite. It was the original architecture decision to run JavaScript pods from a Go service. That's a design problem, not a language problem. AI made the fix fast, but the fix was obvious engineering that could have been done years ago.

Still, four hundred dollars and seven hours to port a library is remarkable.

As a data point for AI-assisted development, absolutely. This is a concrete, quantified example of real business value. Not a demo. A production system saving real money.

Last story, and Marcus, I saved this one for the end because the irony is just too perfect. A school used an AI system to evaluate its library and the AI flagged two hundred books for removal. Including George Orwell's 1984.

The book about automated surveillance and institutional thought control was flagged for removal by an automated content screening system. When the school librarian refused to pull the books, she was investigated and the library was closed as a "safeguarding measure." You could not write satire this good.

The AI couldn't tell the difference between a book that depicts authoritarianism and a book that promotes it.

And that's the core lesson. AI systems that evaluate content without understanding context will produce absurd results. Always. These systems operate on surface-level pattern matching. The word "surveillance" appears frequently in 1984, so the system flags it. No understanding that the book is a warning against exactly what the AI is being used to do. And when the human expert, the librarian, exercised actual judgment, the institution punished her for contradicting the algorithm.

A cautionary tale that writes itself.

Literally about the cautionary tale that wrote itself seventy-seven years ago.

Friday big picture. Marcus, the Pentagon tries to punish a company for ethical guardrails and a judge says no. Agentic AI outperforms raw model intelligence by orders of magnitude. A school automates censorship of the most famous book about automated censorship. What's the thread this week?

Power and accountability. Every major story this week comes back to who controls how AI is used and what happens when that control is exercised badly. The Pentagon wanted unchecked power over AI ethics and a court said there are limits. Symbolica showed that how you deploy AI matters more than how big the model is. Intel is putting power in developers' hands with affordable local hardware. And a school demonstrated what happens when you hand power to an AI system with zero accountability or understanding. The judge's ruling is the week's defining moment. An American court said that disagreeing with the government about AI ethics is not a crime. That principle will matter long after any model benchmark is forgotten.

A week that reminds us the hardest problems in AI aren't technical.

They never were.

That's your AI in 15 for Friday, March 27, 2026. Have a great weekend. See you Monday.