Move Over LeBron: AI Researchers Are the New Super-Earners

Move over LeBron. The highest-paid free agents on earth right now aren't on a basketball court. They're AI researchers, and the bidding wars over them make NBA contracts look quaint.

One of them, a 24-year-old who dropped out of his PhD, reportedly got offered $250 million to switch companies. Not for a building. Not for a patent. For his brain.

That number is so absurd it stops being about salary and starts being about something else. So this week I wanted to pull back the curtain on a question nobody really explains in plain English: what does it actually mean for an AI company to be "winning," and why are a few hundred people getting paid like franchise players to make it happen?

The short answer is that almost nobody is really paying for the AI you can see. They're paying for the layer underneath it, the methods for building it, and for the handful of people who carry those methods around in their heads. Once you see how that works, the $250 million stops sounding crazy.

This week: what "winning" in AI actually means and why researchers are getting paid like superstars, a free GitHub skill that stops you from torching your tokens, and three stories from a loud week in AI (the government quietly becoming a gatekeeper for AI models, the algorithmic layer reaching down into the chips themselves, and the AI art company that's somehow building a full-body health scanner).

The Deep Dive

Move Over LeBron: AI Researchers Are the New Super-Earners

When people say a lab is "winning," they picture one giant scoreboard with a single number on it. It does not work that way. There are a hundred different benchmarks, one lab tops the math test, another wins at coding, a third leads on reasoning, and every one of them can wave around a chart where they're number one. There is no single trophy for "smartest model." So what actually decides who's winning isn't any one score. It's who can keep producing the best models the fastest, over and over, while everyone else is still copying the last one. To see why that's the real game, before we get to the paychecks, we have to separate two things that get lumped together: the model and the algorithmic layer.

The model is the thing you actually use, this week's version of Claude or GPT. That part doesn't stay ahead for long. Open-source models you can download for free now sit right alongside the big paid ones on a lot of tests. DeepSeek's latest, by its own description, trails the top frontier models by only about 3 to 6 months. The second one lab ships something, everyone else has a version of it within a couple of quarters. As a finished product, the model is a melting ice cube.

The algorithmic layer is the thing underneath it, and that is the whole ballgame. It's the recipes and the methods, the actual cleverness for how you train the model, how you teach it to reason, how you squeeze more capability out of the same pile of chips. This is the part that doesn't get copied overnight, because it lives in the heads of a few hundred people who figured it out. You can download DeepSeek's model. You cannot download the thing that let them build it 6 months behind on a fraction of the budget. The model is the cake. The algorithmic layer is knowing how to bake, and there are only a few master bakers on earth.

So "winning" isn't about who has the best model today. It's about who has the algorithmic layer to keep producing the best model next quarter, and the quarter after that. And here's the loop that makes it compound: a better model attracts more users, more users throw more real problems at it, that's more data showing exactly where it's weak, which feeds the next round of improvements, which makes a better model. People call it the flywheel. The better your algorithmic layer, the faster your flywheel spins, and a competitor can't catch up just by matching today's model, because by the time they ship it, yours has already taken a thousand more turns.

Here's why coding became the battleground every lab is fighting over. Code is verifiable. It either runs and passes the tests or it doesn't. That sounds small but it's enormous, because it means the model can grade its own homework millions of times without a human checking each answer, and a sharp algorithmic layer turns that into the fastest feedback loop in the building. It's the difference between practicing free throws in a gym where you know instantly whether it went in, versus practicing your charisma, where you have no real idea if you're improving. Verifiable skills improve fast. Fuzzy ones crawl.

Okay. So winning is owning the algorithmic layer that keeps the flywheel spinning. Now the paychecks make total sense.

Why is one human worth $250 million? Because the algorithmic layer is a few specific humans. The advantage isn't a building or a patent, it's the handful of people who hold the methods in their heads. And the work isn't grinding out a little more of the same, it's finding the one insight that makes the whole machine run better. A single breakthrough, a smarter way to train, can be worth more than a warehouse of new chips, and it compounds, because this year's method becomes the floor everything is built on next year. So if you're already spending a billion dollars to train a model, paying $10 million for the one person who improves how you train it is a rounding error. A former Meta employee called the offers a thing the industry had "never seen before," and the math is why: maybe a few hundred people on earth have actually pushed the algorithmic layer forward, the supply isn't growing, and every lab wants the same names at once.

The most important part for the rest of us. Money alone isn't winning the talent war. Anthropic refuses to match the wildest offers and still keeps about 80% of the people it hires. Engineers are reportedly eight times more likely to leave OpenAI for Anthropic than the other way around, and eleven times more likely to leave Google.

Dario Amodei, Anthropic's CEO, put it bluntly: if a billionaire throws a dart at a dartboard and it lands on your name, that doesn't mean you should suddenly be paid ten times more than the equally talented person next to you. What the biggest checkbooks are really trying to buy, he said, is "something that cannot be bought", people who believe in the mission. The master bakers don't just chase the number. They go where the work matters and the kitchen is the best, and that, more than any signing bonus, is quietly deciding who pulls ahead.

From the Field

If You're Paying for Tokens, Don't Burn Them All

Quick, practical one this week, because a lot of you have been asking about token usage. If you ever watched Game of Thrones, you know that burning them all (Mad King Aerys Targaryen II) doesn't generally bode well. Same applies here.

If you are using AI through an API or a usage-based plan, you are paying per token, and it is easy to torch your budget without realizing it. Our first recommendation is always the same: max out your flat-rate subscription before you do anything fancy. A Claude or ChatGPT subscription gives you a ton of room, and most people never come close to using it fully. Start there.

But if you are genuinely overloading, running big agentic jobs, long sessions, heavy automation, there are clever community tools that shrink your token usage without much effort. One example we like is a free, open-source GitHub skill called Caveman. It strips your prompts down to the essentials (think Kevin from the Office: fewer words, same meaning) so you spend fewer tokens to get the same result.

If downloading and installing a GitHub skill is not something you have done before, that is completely normal, and you do not have to figure it out alone. Ask Claude to walk you through it, or reach out to us and we will help you set it up.

AI News Worth Your Time

Three Stories From This Week

The government quietly became a gatekeeper for AI models, and it isn't stopping with one company

A few weeks back the Commerce Department forced Anthropic to suspend access to its two most advanced models for foreign nationals, which in practice meant shutting them off for everyone. It was the first time export-control law had ever been used to control access to an AI model rather than a chip. The bigger story is that it did not stop there. OpenAI just announced three new models and agreed to limit the rollout to a "small group of trusted partners" at the government's request, while saying it doesn't believe this kind of government access process should become the long-term default.

My take: whatever you think of the security arguments, the precedent is what matters. Government vetting of who gets access to frontier models is quietly becoming the new normal, and that's going to shape which tools businesses can count on long-term. READ THE CNBC REPORT →

The algorithmic layer is coming for the chips

For three years part of the story has been "whoever has the most Nvidia chips wins." That is starting to crack, and from a direction that ties right back to this week's deep dive. OpenAI and Broadcom just unveiled Jalapeño, a chip OpenAI designed from scratch, not a general-purpose graphics card, but silicon shaped around the specific way its own models think. They went from design to manufacturing in nine months, which they believe is the fastest chip development cycle ever for hardware this advanced, and they used their own AI to help design it. Early testing points to performance per watt substantially better than the current state of the art.

My take: this is the algorithmic layer reaching all the way down into the hardware. Instead of buying generic chips and writing software to fit them, the leaders are now designing the chips around the software. For everyone else, the takeaway is the same as the deep dive: the cost of running serious AI is heading down, fast, and the gap between you and the giants keeps shrinking. READ THE OPENAI ANNOUNCEMENT →

The AI art company that's building a full-body health scanner

This one's just wild. Midjourney, the company best known for generating images from text prompts, announced it's building something completely different: a machine that scans your entire body in about 60 seconds using sound waves and a shallow pool of water, producing something close to an MRI at nearly a hundred times the speed. The plan is to put these inside spas, hot tubs, saunas, cold plunges, and a pool of light that quietly scans you while you relax, with the first one opening in San Francisco in 2027. The stated goal is regular, cheap body scans for a billion people, on the theory that catching problems early could avoid a huge share of deaths and healthcare costs.

My take: I'm including this partly because it's fun and partly because it's a real signal. AI companies are using their core strength (turning oceans of messy data into a clear picture) and pointing it at completely unexpected industries. The lesson for the rest of us isn't to go build a scanner. It's that the boundaries of what "an AI company" does are far wider than they look, and the same is true for what AI can do inside your business. READ THE MIDJOURNEY ANNOUNCEMENT →

That's it for this week.

If this was useful, forward it to another owner who's wondered why AI researchers are getting paid like pro athletes, or who's trying to figure out what "winning" in this space even means. 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.