Last week, Google DeepMind dropped something that should have made the world choke on its coffee.
Their new system, AlphaProof Nexus, cracked nine open mathematical problems—real, decades-old monsters that had outlasted careers. Two of them had been sitting there untouched since 1970. Fifty-six years of human genius, and the bill came to a few hundred bucks in compute.
Think about that. A mystery older than most CEOs, solved by an algorithm for the price of a cheap laptop. If this had happened in 2024, the New York Times would have run a special edition. LinkedIn would have melted. We’d all be updating our doomsday decks.
But last week? Crickets. You probably scrolled right past it. I almost did too.
We’re not jaded. We’re exhausted. Breakthrough Fatigue is real. AI is moving so fast, breaking so many "impossible" barriers, that the extraordinary now feels like Tuesday. We’ve become numb to miracles.
The Day the Tests Died
Look at the last two years.
In 2021, MMLU was the gold standard—the SAT for machines. Today, every frontier model scores above 90%. When the entire class gets an A+, the test stops telling you anything useful.
Then came GPQA Diamond. They designed this thing specifically to be un-Googleable. To even qualify as a question, it had to be so viciously specific that only a PhD in that exact sub-field could solve it; a PhD from a neighboring field with full internet access would still bomb.
GPT-4 scored 39%. Respectable, but human.
By early 2026, Gemini 3.1 Pro hit 94.1%. Human PhDs average around 65%. In two years, we went from "worse than a grad student" to "embarrassing the domain experts."
The people writing the exams can’t write them fast enough anymore.
Welcome to the Era of Proof Abundance
Last month, Terence Tao—who is to mathematics what Mozart was to music—stood up at Stanford and said something that shook me.
We’ve left the era of Proof Scarcity and entered the era of Proof Abundance.
It used to be that a major proof was a generational event. Mathematicians would burn their lives down, filling sacks with scratch paper, just to move one conjecture from "maybe" to "true." It was sacred. It was scarce.
Now? The Erdős Problem website has a backlog of over twenty AI-generated proofs just sitting there, waiting for humans to verify. The machines are outpacing our ability to even read them.
Tao admitted he’s hit pause. He can’t keep up. And he used an analogy that’s going to stick with me forever:
AI is like a helicopter that drops you at the summit. You get the view instantly. But you miss the climb. And here’s the thing—the climb is where the value lives.
The Corporate Dilemma: How Do You Measure "AI Proficiency"?
So what does this have to do with your Monday morning standup?
Everything.
In math, the proof is binary. You either cracked it or you didn’t. Objective truth. Beautiful.
But in your office? If you ask AI to draft a marketing plan, build a slide deck, or write a Python script, the output is always... pretty good. It’s never embarrassing. It’s always plausible.
So everyone on LinkedIn claims they’ve "10x’d their productivity." But a CEO pulled me aside last week and asked the question that nobody wants to ask out loud:
"James, my team is using AI everywhere. My API bills are through the roof. But how do I know who’s actually mastering this thing, and who’s just really good at looking busy?"
It’s a killer question. Because without a real way to measure this, we’re all just geniuses in our own Slack channels.
The answer isn’t a tool. It’s a mindset.
Accumulation.
Horizontal Consumption vs. Vertical Accumulation
Watch how your team uses AI, and you’ll see two species emerge.
The Consumer (Horizontal)
They dump their bullet points into ChatGPT. It spits out a polished report. They send it. They saved twenty minutes. They "used" AI.
But ask them what they learned. Ask them what they can do today that they couldn’t do last month. You’ll get a blank stare. They built a sandcastle. The tide came in. Nothing stuck.
The Accumulator (Vertical)
They also use AI to write that report. But then they spend ten extra minutes in the chat. "Look at what I accomplished this week. What’s one technical skill or strategic framework I just used that I didn’t have in my toolkit 90 days ago?"
They log it. They map it. They own it.
Three months later, the Consumer is still offloading busywork to a machine. The Accumulator is visibly different. They can point to specific capabilities they’ve built. They’re not just faster—they’re taller.
So ask yourself: Are you pouring concrete, or are you building sandcastles?
The 2.5% That Matters
Back to AlphaProof Nexus. Nine solved problems. Sounds incredible.
But DeepMind also told us it attempted 353.
Its success rate was 2.5%.
In any other context, that’s a failing grade. But in mathematics, that 2.5% is permanent. Every confirmed proof becomes a foundation. The machine stands on it to reach higher. It doesn’t forget. It doesn’t start over. It accumulates.
That’s the whole game.
You’re not competing with the AI on raw intelligence. You never will. Your edge is your deep, human domain expertise—the stuff algorithms can’t touch (for now). Reading a client’s micro-expression during a negotiation. Sensing when a team is about to break. Knowing which risk to take when the data is 50/50.
You use that judgment to give the AI precise, powerful instructions. The AI gives you leverage. You use that leverage to sharpen your judgment further.
That’s the flywheel. That’s accumulation.
When you build vertically, the AI becomes your engine. When you drift horizontally, it becomes your crutch.
Which one are you building?
JamesCEO, Mercury Technology SolutionsAccelerate Digitality.
