A company spent half a billion dollars on Claude last month. Not on a data center. Not on a moonshot. On licenses, because nobody told employees to slow down, and apparently a few of them were using a frontier model to check the weather.

I would like to humbly point out that this news came out the same week I sent a newsletter called How to Save $14,800 a Month, where the whole point was that a flat subscription beats paying raw API costs for almost everyone. So clearly they read it, took careful notes, and then decided to set half a billion dollars on fire anyway. (They did not read it. The story is real. Half a billion, in a month.)

Here is what I find fascinating about that story. Some people read it and think, holy cow, how do you spend $500 million on a tool that still gets my fourth grader's homework wrong. Other people read it and think, well, yeah, I get it, that same tool just solved math proofs nobody on Earth could crack. Both reactions are completely reasonable. Both people are looking at the exact same technology. That gap is the whole thing I want to talk about today.

This week: the Jagged Frontier and why a vending machine is easier to trust than a genius, the Canva trick quietly giving designers their afternoons back, and three stories from a big week in AI (including a model cracking an 80-year-old math problem, and Opus 4.8 landing while something bigger waits behind it).

Here is something strange about working with AI. The same tool that can pass the bar exam, write working code in seconds, and translate a language almost nobody speaks will then confidently tell you that "strawberry" has two R's. It solves the hard thing and fumbles the easy thing, and it does both in the exact same calm, certain voice.

Researchers at Harvard Business School have a name for this. They call it the Jagged Frontier. The idea is that AI's abilities do not form a smooth line from "easy" to "hard." They look like a mountain range. Tall peaks of genius sitting right next to deep valleys of cluelessness, and the edge between them is invisible.

Normal technology does not behave this way. If a calculator can handle rocket math, you safely assume it can add 2 and 2. Capability scales smoothly. With AI, that assumption is a trap, because the model is predicting likely patterns, not running airtight logic. So it can be brilliant at the thing you expected to be hard and hopeless at the thing you expected to be trivial.

That gap is where people get hurt. Because the AI sounds equally sure of itself everywhere, you start trusting it everywhere and not thinking critically about its capabilities. The moment you wander off a peak and into a valley without noticing, the model keeps talking in that same confident voice, and the mistake sails right past you. Researchers call this automation bias. I call it the reason people end up sending a 30-page report nobody asked for and trusting a number that was quietly made up.

Now compare that to a vending machine. A vending machine has the clearest boundary in the world. You put in money, you press E4, chips come out. If it has the strength to drop a heavy soda, you know it can drop a light bag of pretzels. It will never surprise you by dispensing a flat-screen TV (too bad) and then failing on a pack of gum. And when it fails, it fails honestly. It does not hallucinate a snack or gaslight you into thinking you already ate it. It just holds your chips hostage on a metal coil, and you know exactly whose fault it is.

AI is the opposite. You can hand it a brilliant prompt and still have to cross your fingers it does not trip over a basic step. That unpredictability is the entire reason two smart people can have completely opposite experiences with the same tool.

Now, one important thing, because I do not want to leave the wrong impression. AI does get things wrong, but it is getting things wrong far less than it used to. The valleys are shrinking fast. The new Opus model in the news below is roughly four times less likely to let a flaw in its own code slip by than the version from a month ago. A year ago, hallucinated facts were everywhere. Today they are rarer, and the best models will increasingly tell you when they are not sure instead of bluffing. The frontier is still jagged. It is just less jagged every few months.

Which is exactly why the real skill here is not "catching the AI when it lies." It is knowing what the tool is actually good at. Becoming AI literate. The people who get the most out of AI are not the ones who trust it blindly, and they are not the ones who dismiss it after one bad answer. They are the ones who have a feel for the peaks and the valleys, and aim it at the right work.

So how do you build that feel? Two shifts.

First, stop judging AI at the job level and start judging it at the task level. "Is AI good at marketing" is the wrong question. It depends entirely on how you use it. Ask it to write you twenty subject lines and it will fly. Ask it to go grab coffee with your most important prospect and read the room, and obviously it cannot, and someone watching will walk away saying AI is useless. The in-between is where it bites you. Back in 2023, a New York lawyer asked ChatGPT for case law to cite in a federal filing, and it handed him a stack of confident, detailed, completely made-up court cases. He asked it to double-check. It told him they were real. They were not. He got sanctioned, and the story went around the world as a cautionary tale. You would think that would be the end of it. It was not.

Just this month, an Oregon federal court hit two lawyers with a $110,000 penalty, the largest of its kind yet, for filing a brief full of fake AI-generated cases and quotes. Here is where it gets almost too on the nose. Fortune reported the lawyers submitted 23 fabricated citations and eight invented quotations. But the court record and the legal trade coverage put it at 15 fake citations and eight quotes. Same case, two numbers, in an article that is about AI making up details in legal documents. I cannot tell you for certain why one said 23 and one 15. Maybe a person transposed it, maybe a model wrote that sentence and nobody checked. That is exactly the point. From the outside an AI cannot tell a confident wrong number from a confident right one.

Second, pick how you work with it. Researchers describe two styles. Centaurs split the work cleanly. Human does the human part, AI does the AI part, clear handoff. Cyborgs blend, weaving their own thinking in and out of the AI's in real time, catching things as they go. Neither is wrong. But both require you to actually know where the edges are, which is the part most people skip.

The math proof in the news section below is a perfect peak. The lawyer sanctions are a perfect valley. Same technology. The only thing that changed was whether a human knew which one they were standing on.

If your team is getting wildly different results from the same AI tools, this is almost always why. Mapping where AI is strong and weak for your specific workflows, and getting your people genuinely AI literate, is a huge part of what we do on day one of an engagement.

This one is for anyone who touches design. Marketers, founders building their own decks, anyone who lives in Canva.

You know the work I mean. You spend hours meticulously perfecting a design. Nudging a headline three pixels left. Swapping a color so it matches the brand. Resizing the same graphic four times for four different channels. Copying a line of copy out of one tab and pasting it into another. You have AI open in one tab and Canva open in another, two great tools sitting right next to each other, completely disconnected, and you are the human glue holding them together.

The Canva connector closes that gap. You connect Canva to your AI once, and now the AI can build the design itself. It creates the presentation, fills your branded template, pulls real assets out of your Canva library, resizes one graphic into a LinkedIn banner and an Instagram square and a Pinterest pin, and exports the whole thing. All from the conversation. You never leave the chat to go push pixels by hand.

The honest version: it shines on a paid Canva plan, where your Brand Kit and templates are already set up, so the output comes back on-brand instead of needing cleanup. When the brand setup is right, the time saved compounds fast across every design you make. That is the whole point. It is not one magic graphic. It is the tenth and the fiftieth graphic that each used to cost you twenty minutes of fiddling and now cost you almost nothing.

If you want a hand getting your Brand Kit set up so the connector actually produces work you can use, reply and I will point you the right way.

Opus 4.8 is here, and something bigger is waiting behind it

On May 28, Anthropic shipped Claude Opus 4.8 at the same price as the model before it. The pitch is not "smarter at everything." It is "more honest." Early testers say it flags its own uncertainty more often and makes fewer claims it cannot back up, and Anthropic says it is about four times less likely to let a flaw in its own code slip by unnoticed. There is also a new effort control that lets you dial how hard, and how expensively, the model works on a task (useful, given the $500 million). The more interesting part is what is parked behind it: Mythos, a model Anthropic says sits a full tier above Opus, currently locked to a small group for cybersecurity work, expected to open up "in the coming weeks." Same week, Anthropic raised $65 billion at a $965 billion valuation.

My take: Anthropic called 4.8 "a modest but tangible improvement," which is refreshingly un-hyped for this industry. And honestly, the honesty upgrade matters more than another benchmark point. Go back to the Jagged Frontier piece above. A model that tells you when it is standing in a valley instead of bluffing through it is exactly what makes AI safer to actually put to work in a business. The benchmarks are below. Exciting, not mind-blowing. Mythos is the one I am watching.

READ THE REPORT →

An AI just cracked an 80-year-old math problem humans could not

In 1946, the mathematician Paul Erdős posed a deceptively simple geometry question about how many pairs of dots on a page can be the exact same distance apart. For nearly eighty years, the best mathematicians in the world tried to prove his guess was right. This month, an AI model went the other way and disproved it, finding an arrangement nobody had found before. Real mathematicians, including a Fields Medalist, checked the work and were genuinely impressed. What stood out to them was not raw intelligence. It was patience. Humans mostly tried to prove Erdős right because that felt likely. The AI was willing to grind down a tedious, unpromising path no human would have bothered to finish, and that is exactly where the answer was hiding.

My take: I do not think math is where AI is going to run away from us fastest. It is a narrow, checkable, very specific kind of work. But this is a clean example of a peak on that jagged frontier, and it teaches one real lesson for the rest of us. Put a genuinely capable tool in the hands of genuinely capable people and you get results neither could reach alone. The AI did not replace the mathematicians. It went where they would not, and they confirmed where it landed. That is the Centaur move, at the highest level there is.

READ THE SCIENTIFIC AMERICAN STORY →

The AI spending hangover is starting

The $500 million Claude bill from the top of this email was not a one-off. It is the leading edge of a broader correction. Companies that went all-in on AI are now asking hard questions about whether the spend is producing real results, and some are cutting back. Amazon is shutting down its internal "tokenmaxxing" culture, with a senior exec telling staff to stop using AI "just for the sake of using AI" and to point it at real business problems instead.

My take: This is the whole Bridge AI thesis showing up in a headline. The tool is not the win. We have said it in basically every issue. Spending more on AI is not the same as getting more from AI, and the companies learning that the hard way are the ones who bought the licenses and skipped the implementation. Buying the gym membership is not the same as getting in shape. The work is in what your team actually does with it, on the tasks that actually matter, with a human who knows where the frontier's edges are.

READ THE FAST COMPANY STORY →

That's it for this week.

If you found this useful, forward it to someone who keeps having the same argument about whether AI is brilliant or useless. They are probably both right, just standing on different parts of the frontier.

And if you want to talk about what AI could actually do for your business, not in theory, but in your workflows, with your team, on your timeline, just reply to this email. That's what we're here for.