Scrambling Up The Stack
Last week, I wrote about building a dry-humor travel website with Claude Code — how Claude abstracted away all the busywork and let me focus on just what mattered: the conception and the voice. Cristi looked at what I built and said, "this is not software, this is art." Which was insanely generous, but had a kernel of truth. AI tools can take away all the busywork and let humans focus on the art, the craft, the creativity, the ideation, the frontier.
I've bumped across this idea in different ways this week.
I met with Chris Jones this week of Surfboard. I've always enjoyed working with Chris (he reminded me we started working together in 1996!!), and I love what he and his team are doing at Surfboard. They are building the indispensable AI-based productivity tool that works across your entire business and lets you get things done. Surfboard connects to your systems, knows your business, and lets you focus on the high-value add decisions on top of all that, rather than futzing around with documents, reports, spreadsheets, dashboards, and whatever.
Once upon a time, I viewed Microsoft Office as the indispensable software that I had to install first thing on a new PC. Now I've let my Microsoft 365 license lapse because I no longer need to author spreadsheets, docs, and presentations directly. I'll let Surfboard or Claude output whatever doc formats they need to output for the task at hand.
Surfboard isn't the only company chasing this dream. Claude Cowork, some instantiation of Microsoft Copilot, and many others will be in the space. I like Surfboard because it is approachable and simple to get going with, and not locked behind some complicated enterprise license scheme.
Derek Thompson wrote this week:
This is what the history of multiple discovery is actually telling us. The great bottleneck of progress is question-framing. Once a problem is framed with sufficient clarity and precision, the answer almost wants to be found. Once Malthus articulated his grim theory of resource scarcity and competition precisely enough, two scientists on opposite sides of the world arrived at the same revolutionary solution within years of each other. The answer was, in some sense, already waiting.
Multiple discovery arises because, at some point, the question is well-framed, allowing scientists to wrestle with the real issue and find the solution. It is the framing that is the bottleneck.
Our AI tools can execute like crazy; they can out-execute us feeble humans — they are diligent, untiring, exacting. What they don't do well is framing the right problem, finding the right voice, and conceiving the thing worth building. AI will get better at problem-framing, but it is an unending challenge.
I talk to engineers every day who use AI to write their code.
They all say the same thing: AI gets them 95% there. Fast. Sometimes shockingly fast.
But that last 5% is where the real engineering happens. The edge cases. The architecture decisions. The stuff AI doesn't know it got wrong.
Karpathy calls this the "march of 9s."
Getting from 90% to 95% correct is easy. 95% to 99% is hard. 99% to 99.9% is where you need someone who actually understands what's happening underneath.
Every additional nine costs exponentially more judgment.
AI gets you to 95% solutions fast — sometimes shockingly fast. But the last 5% is where the real work is. The edge cases, the architecture decisions, the things AI doesn't know it got wrong. Every additional nine costs exponentially more judgment – and there is always another nine.
As everyone notes, AI eats the bottom of the stack — technical execution, productivity busywork, wide swathes of effort that were never really about ideas in the first place. And as it does, the bottleneck doesn't disappear. It moves up, towards greater judgment and better framing.