The AI War Nobody Voted On Just Got Real.
Plus: A new $599 laptop built for AI, the tool that turns any document into a cinematic video, and why the biggest winners in journalism right now are the ones using AI — not fighting it.
I’ve been doing this newsletter long enough to know when a week is genuinely different.
Most weeks, the AI news is impressive but safe. New model, benchmark scores, product launch. You read it, you file it away, maybe you try something new.
This week was not that week.
This week, the company behind the AI tool I use every day got officially labeled a national security threat by the U.S. Department of Defense. A data center running military AI got bombed. And the two most powerful AI labs in the world are now in open warfare — not just over customers, but over what these tools should be allowed to do at all.
I’m not going to pretend that’s normal. It isn’t.
Word of the Day: AI Guardrails
Guardrails are the rules and restrictions built into an AI system to prevent it from doing harmful things.
Think of them like the bumpers at a kids’ bowling alley. They don’t make the ball go faster. They keep it from going somewhere dangerous.
For your business, this matters because guardrails determine what your AI tools will and won’t do. They’re also, apparently, worth going to war over.
The Biggest Story in AI Right Now: The Claude Exit Tax
I need to give you the timeline on this one, because it’s moving fast — and because it may affect your business more directly than you think.
Here’s what happened, in plain English:
Back in February, the U.S. Secretary of Defense gave Anthropic — the company that makes Claude — an ultimatum. Give the military unrestricted access to your AI for “all legal purposes” by Friday, or lose your $200 million defense contract. And get blacklisted from working with anyone else in the government supply chain.
Anthropic said no. Specifically, they said two things should never be allowed: mass domestic surveillance, and fully autonomous weapons. They published that position publicly.
The Pentagon responded by officially labeling Anthropic a “supply chain risk to national security.” Effective immediately.
Then this week, Anthropic and the Pentagon sat back down to try to reach a deal. They got close. One sticking point: a phrase about “bulk data analysis.” The negotiations collapsed again. Anthropic is now planning to challenge the designation in court, arguing the law was designed for foreign threats like Huawei, not American companies.
Meanwhile: OpenAI swooped in and signed its own deal with the Pentagon. Sam Altman publicly criticized Anthropic, saying private companies shouldn’t override government authority on national security.
And in the background, a detail that will keep you up at night: Iran reportedly bombed AWS data centers in Bahrain that were being used to run military AI workloads.
Here’s where it gets personal for your business.
Tech analyst Shelly Palmer gave this situation exactly the right name: The Claude Exit Tax.
Here’s what that means. If your company does any work with the U.S. government, the military, or any company that does — you now have a compliance problem if you’re using Claude. The “supply chain risk” label is historically reserved for foreign adversaries like Huawei. It means any organization connected to DoD contracts has to think hard about whether Claude is on their approved vendor list.
If you’re a government contractor, a defense subcontractor, or you sell to agencies that are — this is not “watch this space” territory. This is “talk to your compliance team this week” territory.
If your business has zero government exposure, you’re fine for now. But the rules are being written in real time, and they’re moving faster than anyone expected.
The New Model That Beat Humans at Their Own Jobs
OpenAI didn’t slow down while all that drama was unfolding.
They released GPT-5.4, and it’s genuinely impressive.
Here’s the headline number: on a benchmark that tests real desktop navigation — actually clicking around a computer, finding files, completing tasks — GPT-5.4 scored 75%. The average human scores 72.4%.
The AI is better than most people at navigating a desktop computer.
It was also tested across 44 real job categories. It outperformed or matched professionals 83% of the time.
Other upgrades worth knowing:
A 1 million token context window. In plain English: you can now feed this model an enormous amount of information in one shot. We’re talking an entire book, a year’s worth of emails, or a company’s complete document library. It reads all of it at once.
A new extreme reasoning mode that essentially tells the AI to slow down and think harder before answering. Better results on complex problems. Slower, but more accurate.
OpenAI researcher Noam Brown summarized their position this week: “We see no wall.”
What this means for you: GPT-5.4 Thinking is available now for Plus, Team, and Pro users. If you have a subscription, it’s already in your account. Try it on your most complex, multi-step business task this week. You may be surprised.
The Tip of the Week: Turn Any Document Into a Cinematic Video in Minutes
This one genuinely made me do a double-take.
Google’s NotebookLM — which you already know can turn documents into AI podcast audio — just added something new: Cinematic Video Overviews.
You upload your documents. The AI reads everything. Then, instead of just summarizing, it produces a fully animated, narrated, studio-quality video explaining the content.
We’re not talking a PowerPoint with voiceover. We’re talking dynamic visuals, transitions, and a narrative arc — built automatically from your source material.
It uses a combination of Google’s Gemini 3 model as the “director” (handling the story structure and visual consistency), plus their video generation model.
Real use cases for business owners:
Upload your Q1 report, your customer survey results, or your onboarding documents — and get a watchable video version in minutes. Share it with your team, your board, or a client who’s never going to read a 40-page PDF.
The honest catch: Right now it’s only available for Google AI Ultra subscribers (the paid tier), English only, and capped at 20 generations per day. It’s not a free feature.
But the direction is clear. The era of “nobody read my report” is ending.
NotebookLM is at notebooklm.google.com. You can explore the free features there and see the pricing for Ultra.
The Story That Media People Are Quietly Obsessing Over
There’s a debate playing out in newsrooms across the country right now. And honestly, it applies directly to your business.
The question: Should you use AI to create content?
The Philadelphia Inquirer is using AI tools to scan community meetings, flag stories worth covering, and help produce local newsletters for areas of Pennsylvania and New Jersey they never had the staff to cover before. Those newsletters have attracted more than 50,000 free subscribers in under a year. They’re calling it a subscription driver.
The argument from the people doing it: If it’s the difference between covering an underserved community and not covering them at all, why wouldn’t you?
The argument against: AI can mimic but it can’t truly understand. It doesn’t have judgment, lived experience, or ethical instincts. A reporter who shows up and listens is irreplaceable.
Both are true.
For your business: This same tension lives in your marketing, your customer service emails, your social posts. AI can produce volume. You provide the judgment, the standards, and the relationships. The winners right now are the ones who figured out which work belongs to each.
News Corp this week signed a $50 million per year deal with Meta to license their content for AI training. Their CEO summed up their strategy as “woo and sue.” They’ll work with you if you pay them. They’ll take you to court if you steal their content.
Smaller publishers — and smaller businesses — don’t have that leverage. Which means the rest of us have to get smarter about how we use AI, and smarter about protecting what we create.
Quick Hits Worth Knowing
Apple launched the MacBook Neo. Starting at $599, built on the A18 Pro chip, with full Apple Intelligence features built in. It ships March 11. This is the most affordable Apple laptop ever with real AI capabilities built in from day one. If you’ve been waiting to get into that ecosystem, this is the entry point.
Google launched a Workspace CLI. A free tool that lets AI agents directly read your emails, schedule calendar events, edit documents, and manage files in Gmail, Drive, and Sheets automatically. This could replace a lot of paid automation software. Worth watching if you use Google Workspace.
xAI (Elon Musk’s AI company) is building 1.2 gigawatts of computing power. For context, that’s more electricity than many mid-sized cities use. The AI infrastructure arms race is not slowing down.
Alibaba released four new open-source AI models. Ranging from tiny (great for phones and local devices) to mid-size, all free to use and among the best in their size category. If you have a developer on staff, these are worth knowing about.
A woman in New Zealand built a public website using Claude that shows Auckland ferry timetables in a readable format. The original was buried in a PDF on a government site. Her AI-built site got 5,000 organic visitors in less than a week. That’s Vibe Coding. That’s the whole thing, right there.
The Bottom Line
Here’s the honest truth about a week like this.
The tools are better than ever. GPT-5.4 is legitimately impressive. NotebookLM is doing things that would have required a full production team two years ago. A woman in New Zealand solved a real public problem with a couple hours of prompting and a free tool.
At the same time, the company that makes the AI I recommend most is now in a legal fight with the U.S. military over what that AI is allowed to do. A data center running military AI workloads got bombed. News Corp is getting $50 million a year from Meta just to not sue them.
None of that is normal.
I don’t say this to scare you. I say it because you’re a soon-to-be AI leader, and leaders who pay attention win. The people who understood the internet in 1998 built things that lasted. The people who figured out mobile in 2010 got ahead. Right now, the window is wide open.
But the window doesn’t stay open.
Try GPT-5.4 Thinking this weekend on your most complex problem. That’s your homework.
SmartOwner is published (almost) daily by the team at DigitalTreehouse. Want AI consulting or automations for your business? Reply to this email.