Who Gets to the Moon First — You or Your Competitor?

Over the holidays, something important happened — and it wasn’t a model release.

A lot of the people closest to the frontier finally had what they rarely get: unstructured time. No meetings. No roadmaps. No demos. Just room to tinker.

And what came out of that “play time” wasn’t a bunch of cute experiments.

It was a realization that made even experienced builders uneasy.

Andrej Karpathy captured it perfectly:

“I’ve never felt this much behind as a programmer… I have a sense that I could be 10X more powerful if I just properly string together what has become available.”

Read that again. That’s not a casual observer. That’s someone who helped build the modern AI era.

If people at that level feel the gap, here’s the uncomfortable implication:

Capability is compounding faster than institutional learning.

So when people ask, “What should we do with AI?”, I think they’re starting in the wrong place.

The question that matters is:

Who gets to the moon first — you or your competitor?

Because we’re moving from AI that answers to AI that executes. And once execution can run while you sleep, the clock speed of the business changes.

I tell this to my friends all the time — Do you want to win?

The signal behind the headlines

Cursor recently published a research write-up describing what happens when you stop treating AI as a chat window and start treating it like a long-running execution system. (I wrote a short post about this on LinkedIn a few days ago)

To stress-test their approach, they picked a deliberately hard goal: build a web browser from scratch. Their report describes coordinating hundreds of concurrent agents — with distinct roles (planners, workers, and a judge) — running for close to a week, producing over a million lines of code across ~1,000 files, and consuming trillions of tokens.

Their CEO described a related run using GPT-5.2 inside Cursor for a week that produced “3M+ lines of code,” with a browser that “kind of works” and can render simple websites “largely correctly.”

Argue about whether it’s a “real browser” if you want.

The browser isn’t the point. Multi-day, multi-agent execution is the point.

Once that’s real, the constraint changes:

• Building gets cheaper
• Iteration accelerates
• The bottleneck becomes verification, governance, and deployment discipline

Which is where most organizations are exposed — because they’ve been optimizing for generation, not operations.

Sierra is a good template to understand: autonomy that improves every day

This is why Sierra is such a useful example — not because customer service is the only use case, but because it shows the loop every industry is going to need.

Two points that should land with any executive:

• Most of the cost of software is maintaining it, not building it.
• We’re over-focused on “vibe coding” the old world of screens and dashboards, when the future interface is agents.

If you take that seriously, you realize what the race is really about:

Not who can generate software. Who can operate autonomy.

Sierra’s posture is essentially an “agent operations” loop: deploy → monitor → identify failure modes → improve → redeploy. Over time, the system becomes more robust because it’s learning from the real world.

That’s the difference between an AI demo and an AI advantage. Think exponential improvement.

The inversion leaders are missing

As generation becomes cheap, trust becomes expensive.

So, the moat shifts to “Mission Control”:

• Acceptance criteria
• Testing
• Permissions
• Audit trails
• Rollback
• Integration discipline

“Not production-ready” isn’t a dismissal. It’s a timeline marker.

The organizations that build Mission Control will compound. The organizations that treat this like a gadget will create chaos — and then retreat.

Governance is now existential

Once agents can take actions that affect external state — files, systems, money movement, customer accounts — the security conversation stops being optional.

That’s why the National Institute of Standards and Technology (NIST) Center for AI Standards and Innovation issued a Request for Information on securing AI agent systems: the risk profile changes when model outputs are connected to tools that can do real work.

You don’t “adopt agents.”

You design the safety envelope they operate inside.

The catch-up math

If your competitor starts running real autonomous loops six months before you do, they don’t just get six months of output.

They get six months of:

• templates and checklists
• evaluation harnesses
• failure libraries (“don’t do that again”)
• operational muscle memory
• workflow redesign that enables more workflow redesign

That’s compounding.

Can you catch up? Yes — but only if you change your rate of learning.

If you try to catch up with a cautious pilot that never becomes an operating model, you won’t catch up.

You’ll produce nice decks.

The question

If your AI strategy is still “a chatbot” or “a copilot,” you may want to rethink it.

Here’s the question:

Which workflow can you let run overnight — inside guardrails — so humans wake up to a verifiable work product?

If the answer is “none,” you’re not in the race.

Your competitor however probably is.

All that said — It is a damn exciting time to be creating, learning and pushing the envelope.