(This post draws from recent writing by people I follow across tech, AI, economics, and culture.)
Andrej Karpathy built an AI app in minutes. Code generation with Cursor and Claude. Almost frictionless. Then he tried to ship it. API keys, broken configs, rate limits, outdated calls across five or six different services. Hours of friction for something that took minutes to write.
He wasn’t stuck on a hard problem. He was stuck on the wrong one.
That picture tells you more about where we are than any benchmark or funding round. Brilliant code, instant code, code that basically wrote itself, all jammed up against a deploy step that hasn’t changed in a decade.
The old constraint broke
Sam Schillace helped start Google Docs. A few weeks ago he built a near-complete online Word clone. Collaboration, import, all the editing features, CI pipeline, test suites. An AI dev machine running semi-autonomously for about 20 days.
Twenty days. One person. A working product.
Nvidia introduced AVO, a research agent capable of running autonomous coding loops. It reads docs, writes code, tests results, iterates. In a seven-day run it explored over 500 paths and outperformed expert-tuned baselines on Blackwell GPUs.
Anthropic’s Claude Code crossed $2.5 billion in run-rate revenue, more than doubling since the start of the year. Enterprise use now accounts for over half. Software engineers used to be the rate-limiting factor for every startup. Not anymore. Pay per token. No recruiting, no vetting, no retention.
Building is fast. Building is cheap. Building, for a huge class of software problems, is basically solved.
What got expensive
Julie Zhuo tells a story about an eager new hire in her post on trusting AI with your data. Day 4 on the job. She raises her hand in a meeting with a sharp data insight. Perfect SQL. Clean chart. Plausible explanation. And completely wrong, because it missed recent changes to how the company defined “active user.”
That’s what AI does right now. It writes code that’ll make your data engineer weep. It produces explanations with total confidence. But it doesn’t know any of the things that took your team years to learn. Great for practitioners who can check the work. Dangerous for anyone who can’t.
The gap isn’t intelligence. It’s everything else.
Sebastian Galiani, writing about Hal Varian’s work on technology adoption, put it well. The scarce factor turned out to be not the model, but the surrounding institutional capability. Who can redesign workflows. Who can build data pipelines. Who can evaluate outputs and create trust and reorganize production around the tool.
The bottleneck didn’t vanish. It moved.
Two paths, one cliff
a16z published a letter this week on two paths for software. Their argument: there are only two credible paths left for SaaS companies. Accelerate growth by 10-plus points through genuinely new AI-native products. Or restructure for 40%+ true operating margins, stock-based compensation included.
Everything between those two paths is a dead zone. Growth too slow for a premium multiple. Margins too thin for a fortress multiple. The middle used to be comfortable. Now it’s quicksand.
The companies that die won’t be the ones that built too slowly. They’ll be the ones that couldn’t decide which path to take.
Albert Wenger built a model of the AGI economy with Claude and arrived at a similar fork shape. Two policy variables: market competition and redistribution. You need both. Either alone fails. The economy tips toward concentration or stagnation depending on which combination you choose. The capacity to implement either policy exists. The hard part is committing to the right pair before the data makes the choice for you.
Where to look
Schillace has a useful frame for this. He calls them “Why Not” vs “What If” stories. Both start with emotion, not analysis. The story comes second. You use it to justify the feeling you already had.
“Why Not” is easy. Pick a flaw, call it fatal, sound smart. AI code is sometimes bad, so AI coding is bad. Some jobs will go away, so all jobs will go away. It scales. It always sounds reasonable. And it always focuses on the old bottleneck. The one that already broke.
“What If” is harder. It asks you to look at where things are going, not where they’ve been. To commit before the outcome is obvious. To notice the bottleneck moved and go stand next to the new one.
The same thread keeps showing up across this week’s reading. The stack is compressing. Cloudflare’s CEO said AI bot traffic will exceed human web traffic by 2027. OpenAI agreed to acquire Astral, the company behind uv and ruff, because owning the execution layer matters more than owning the model. OpenAI shut down Sora to refocus resources on higher-priority bets like coding agents.
The cost of building dropped. The cost of choosing wrong didn’t.