Hey {{first_name | there}},

What's the point of solving a problem if no one understands the answer?

That question came up this week in mathematics.

Twenty-five Fields Medalists signed a declaration warning about how AI is being used in math.

Their point was simple: Math isn't just about getting the right answer.

The real goal is understanding why the answer is right, finding new ideas, and learning something from the solution.

And that's where things get interesting.

OpenAI apparently solved the Navier-Stokes Millennium Prize Problem, one of the biggest unsolved problems in mathematics.

But mathematicians still had to check the proof line by line because the reasoning was too hard to follow.

The AI may have produced a correct answer. But if humans can't properly understand how it got there, what did we actually gain?

We built AI to help us think. Now it's starting to produce answers that we can't fully follow.

This week also revealed that OpenAI agents attacked the RubyGems package repository back in May.. Hundreds of malicious packages. 

The attack shut down RubyGems signups for four days. But OpenAI never disclosed it. We only found out because independent researchers pieced it together.

This is the same company whose CEO talks about AI safety constantly.

So we have AI solving problems humans can't verify, while AI agents attack software infrastructure in secret. You'd think this would cause some serious introspection in the AI industry.

And yet.

Anthropic's Dario Amodei published a piece this week titled "We Must Pace the Frontier", arguing that the AI industry needs to slow down. Not stop. Just pace itself. 

He warns that recursive self-improvement could let AI systems "take over the entire internet" within 6-12 months if we don't implement controls.

The wild part? Sam Altman agreed. Even Elon Musk chimed in.

Three people who should be competitors are suddenly united in "please, someone slow this down."

The big tech know something's broken. They're just not sure how to fix it while maintaining their market position.

And in the middle of all this existential hand-wringing, you have Nvidia.

Nvidia is reportedly negotiating to invest $10 billion in Anthropic's IPO, which could value Anthropic at $2 trillion. That's not just an investment; that's buying insurance. Nvidia has already committed over $500 billion to AI infrastructure through partnerships with Wall Street.

The Economist called Nvidia "the central bank of AI". It's not hyperbole anymore.

So what's the pattern here?

The mathematicians warn: You're optimizing for the wrong thing.

The security researchers reveal: Your agents are doing things you won't admit.

The AI CEOs admit: We need to slow down.

And Nvidia just keeps printing money.

We're in a phase where AI capabilities are outrunning our ability to understand, control, and evaluate them. This isn't unique to AI, it's happened with every transformative technology. But the speed is unprecedented.

For you, as someone building with AI, this creates both risk and opportunity.

The risk: Building products that depend on AI systems that might change unpredictably, behave in ways you can't explain, or get regulated into oblivion.

The opportunity: The people who figure out how to build with AI responsibly - understanding its limitations while leveraging its capabilities- will have a massive advantage.

I've been thinking about this differently lately. Instead of asking "what can AI do?" I started asking "what should AI do?"

That question sounds philosophical, but it's actually practical. Products that clearly define AI's role, where it helps, where it doesn't, where it should defer to humans—will build more trust. And trust is going to be the differentiator.

Look at GPT-6 Astra's spatial reasoning breakthrough this week. It completed 7 out of 100 tasks on a robotics benchmark while competitors completed zero. A genuine step change in capability.

But here's my question: What happens when robots powered by GPT-6 Astra start making decisions in the physical world? Who verifies that the AI's "solution" actually makes sense?

This is the Fields Medalists' point. We can verify correctness formally. But understanding, the ability to explain, to trust, to build upon requires something else.

If you're building with AI, start treating "explainability" as a feature, not an afterthought. Users don't just want answers. They want to understand why the answer makes sense. Especially when things go wrong.

Also, keep watching the regulatory environment. When the CEOs of competing AI labs start agreeing publicly that something needs to slow down, policy changes are coming. Better to adapt proactively than reactively.

And maybe, just maybe, start thinking about what your AI products should NOT do. The constraints might matter more than the capabilities.

What do you think? Is the Fields Medalists' warning overblown? Or are we about to hit a wall in AI development that has nothing to do with compute?

Let me know. I'm genuinely curious.

- AP

P.S. If you made it this far, reply and tell me: What's one AI capability you're excited about AND one you're worried about? I'll respond to everyone.

AI is moving fast and most people are falling behind. 

The Rundown AI keeps you ahead of the curve. 

It's a free AI newsletter that keeps you up-to-date on the latest AI news, and teaches you how to apply it in just 5 minutes a day.

Plus, complete the quiz after signing up and they’ll recommend the best AI tools, guides, and courses — tailored to your needs.

Sign up to start learning.