You Can't Copy What Doesn't Exist

Every class had that kid. Never did the homework, copied yours in homeroom, and changed just enough words to not get caught. In AI right now, that kid has a research lab, billions in funding, and a spot near the top of the leaderboards.

This week I want to talk about the scariest sentence in AI policy: "China is only months behind." It gets used as the ultimate argument against putting any rules on the big AI labs. Regulate Anthropic, OpenAI, or Google, the thinking goes, and you hand the frontier to China or worse.

I think that fear has it exactly backwards. And once you see how the followers actually stay close, you might too.

This week:

The copycat problem: why "months behind" doesn't mean what you think it means

Claude in Chrome, and the automations hiding behind your logins

Apple takes OpenAI to court, Claude gets its own Spotify Wrapped, and Meta finally ships a good model

The Deep Dive

You Can't Copy What Doesn't Exist

The White House's own AI advisor has said China is only three to six months behind the US in AI. Fact checkers who dug into the question landed in the same place: the gap is measured in months, not years.

That number does a lot of work in the regulation debate. Any time someone proposes guardrails for the big labs, it gets thrown back like a mic drop. Slow down our frontier, and someone else's frontier takes the lead.

Here's my problem with that argument. It assumes the labs chasing the frontier are running their own race. From what I can see, they aren't.

Look at how the followers actually stay close. Meta's original Llama model leaked onto the internet within a week of release, and suddenly everyone on earth had a near-frontier model to study. Other labs release their weights on purpose, and every copycat downstream gets a free blueprint. And then there's distillation, where you ping a frontier model millions of times and train your own model to imitate the answers. That one's not hypothetical. OpenAI has accused DeepSeek of doing exactly that.

Between the leaks, the purposeful weight releases, and the distillation, being "four months behind" starts to look less like a race and more like a photocopy with better margins. The copy can only exist after the original. The follower's speed is set by the leader's speed.

Which flips the whole regulation debate on its head. The fear is that if we regulate the big three, the copycats surge past us. But copycats can't copy what doesn't exist. If the frontier slows down, there's nothing new to imitate, and the whole field slows down together. The four month gap doesn't shrink. It just follows the leader at whatever pace the leader sets.

Think back to the kid who copied your homework. If you didn't do the assignment, he failed too. He didn't magically start doing the work himself the one day you decided not to. Copying was his whole system. And sure, skipping the homework cost you a grade. But anyone who lived this knows there was a quiet satisfaction in watching him turn in a blank page for once.

Now, I'll push back on my own argument. Copying a frontier model well is legitimately hard. DeepSeek's engineers are excellent, and serving a capable model at a fraction of the cost is a real skill that benefits real businesses. I'm not saying open models are useless. I use the fact that they exist as a negotiating chip for clients all the time. I'm saying being a fast follower and setting the pace are two different jobs, and only about three companies on earth are doing the second one.

So when you hear "we can't regulate AI or China wins," ask the obvious question. Wins what? You can't win a race by drafting behind the leader. And if you're trying to figure out which of these models actually belongs in your business, that's a conversation we have with clients every week. Happy to have it with you too.

From the Field

Your Logins Are Not a Dead End

Every week someone asks me a version of "can Claude connect to this?" Their supplier portal. Their state licensing site. Their scheduling tool from 2012. And sometimes the honest answer is no. There's no connector, no integration, nothing to plug in. The data lives behind a login on a website that has never heard of AI.

Here's the move most people miss: Claude in Chrome.

Take a distributor portal. Every Monday, someone on your team logs in, checks the status of 40 open orders, and copies them one by one into a spreadsheet. An hour of clicking, every single week, forever.

With Claude in Chrome, you log into the portal yourself, in your own browser, the same way you always do. Claude never sees or touches your password. Then you set it loose on the part you hate. It clicks through the orders, pulls the statuses, flags the two that slipped, and drops everything into your tracker. The Monday hour becomes about five minutes, and a human still reviews the sheet before anything goes out to a customer.

Why this travels: every business has at least one login like this. If your team signs into a website more than once a week just to copy things out of it, that's your candidate.

AI News Worth Your Time

Three Stories From This Week

Apple takes OpenAI to court

Apple sued OpenAI on July 10, accusing it of systematically stealing hardware trade secrets. The suit names two former Apple employees, including Tang Tan, who spent 24 years at Apple before becoming OpenAI's chief hardware officer, and alleges everything from confidential project code names used in recruiting to a senior engineer who kept an Apple laptop and downloaded technical documents on his way out. OpenAI, which bought Jony Ive's design startup io for $6.5 billion, is widely reported to be building an AI-first device meant to rival the iPhone.

My take: the AI talent wars just became the AI lawsuit wars. When the scarcest resource in tech is the twenty people who know how to build a thing, expect the fights over those people to get uglier. And note what Apple is really telling you here: the biggest company in consumer hardware believes the next iPhone-sized threat is an AI device. READ THE TECHCRUNCH REPORT →

Claude got its own Spotify Wrapped

Anthropic launched Reflect on July 9, a dashboard that works like a mix of Spotify Wrapped and your phone's screen time report. It shows your most active days, your top topics, and how your Claude usage breaks down over the last month up to the last year, and it's available on free accounts, not just paid ones. The clever part: it makes suggestions based on your habits, like recommending Projects if you keep repeating the same context, or custom skills if you run the same task over and over.

My take: open your Reflect tab this week. If every conversation is a one-off question, you're using AI like a search engine, and the dashboard will show you that in one glance. It's the fastest AI audit you'll ever run on yourself, and it's free. READ THE ENGADGET REPORT →

Meta finally ships a good model

Meta released Muse Image, a free image generator from its Superintelligence Labs group, built into the Meta AI app, Instagram, and WhatsApp, with a video model already on the way. Early rankings put it near the top of the community image leaderboards, which is a sentence nobody has written about a Meta model in a long time. The launch wasn't clean, though. A feature that let people restyle other users' public Instagram photos with AI set off a privacy backlash, and Meta pulled it within the week.

My take: after years of Meta models being "fine," real competition at the top is good news for everyone's pricing. But the backlash is the better lesson for business owners. Meta shipped an AI feature that touched people's faces without asking first, and paid for it in days. If you're rolling out AI anywhere near your customers' data, get consent right before launch, not after. READ THE META ANNOUNCEMENT →

That's it for this week.

If this was useful, forward it to the person who keeps telling you nothing can be done about AI because China will win anyway. That's who this is for.

And if you want to talk about what AI could do for your business, not in theory, but in your workflows, with your team, on your timeline, reach out at [email protected]. That's what we're here for.