Anthropic put out a piece recently about how fast their own engineers are moving now. The numbers are the kind that used to sound like marketing: task-completion capability roughly doubling every four months. Over 80% of the code merged into their own production systems is now written by their own model — up from single digits a year ago. Engineers shipping 8x more per quarter than 2024.
Read that twice. That's not a vendor claiming their tool is fast. That's the company that builds the model, talking about its own team.
So naturally the next thought is: if the people building the thing are moving 8x faster, what's left for the rest of us to do?
Here's where it gets interesting, because there's a study that says almost the opposite. METR — an independent group that actually measured developers working with AI tools on real repos, not self-reported — found they were 19% slower, not faster. Same year, same technology, opposite result.
Both are true. They're just measuring different things. METR measured individual developers, mostly without a deliberate process wrapped around the AI's output. Anthropic's number comes from an organization that built code review, merge gates, and engineering ownership around the model before they ever let the 8x happen. The speedup didn't come from the model alone. It came from a human-governed process deciding what gets shipped and what gets sent back — the model just got faster inside that structure.
That's the part that doesn't make it into the "AI writes code now" headlines: the acceleration only shows up where a human is still the one closing the loop.
One line from the same piece, from someone inside Anthropic, on a day the automation was working almost too well: "On days where everything works well, I can't help but think nothing I do matters, everything is automated."
That's a real thing to feel, and it's worth sitting with for a second — because it's the wrong conclusion. The same piece answers it a few paragraphs later, more carefully: "The comparative advantage of humans as of right now is still in seeing the bigger picture."
Not writing the code. Not reviewing every line. Seeing the bigger picture — deciding what's worth building, when it's actually done, and whether to ship it.
That's the exact shape of the gap I keep running into with side projects. The build stopped being the hard part a while ago. Nobody's stuck because the code doesn't compile. People get stuck holding a working app at 80%, because the last 20% — deciding it's real, closing the loop, actually going live — was never a capability problem to begin with. It's a "who's accountable for finishing this" problem, and a faster model doesn't touch that.
Anthropic's own numbers are proof of the same pattern at a much bigger scale: the speed comes from the model, but the bigger picture — what ships, what doesn't — still needs a human holding it. AI tracks. A human reads.
If you've got something sitting at 80% right now: mvpbuilder.io.
Top comments (0)