Karpathy mentioned he hasn't written a line of code since December [0], says agents only actually started working around then. The trajectory in this report lines up with that pretty well: throughput was basically flat for most of 2025, picked up ~25% in Q4, then more than doubled in Q1 2026. I've talked to people at a couple of companies that lean heavily on Claude Code internally, and fixing already-merged code still eats a big chunk of their week. Mostly because nobody's really tracking what's landing in the codebase anymore. So I went into this report expecting that pain to show up as a spike in "fixes." It kinda doesn't though? Fixes only went up 3pp (15→18%), while "growth" work picked up 7pp. So either the cleanup is happening earlier, in code review before merge, or commit data just can't see it. Curious if either of these matches what you're seeing. [0] https://x.com/karpathy/status/2026731645169185220
HN user
Fawlty
FWIW it’s trending on huggingface papers and there is a more in-depth discussion there[1]. Seems like the early results are promising and someone replicated that independently but it still needs to be proven that the claims work also at the bigger scale. [1] https://huggingface.co/papers/2509.26507
Theme Hospital (c $4) and CorsixTH is an amazing combo - they brought back all my childhood memories, highly recommended
It’s cool that you’ve built it! It truly gets complicated, especially when at later stage you start getting venture debt (not safes), more demanding LiqPref and you’re facing an exit at a price which is not ideal. Tbh I usually end up with bespoke excel files for each company to have a full understanding of waterfall for all parties and be able to simulate their payout and incentives at different exit price. But I’m sure even this tool can be helpful - if you need any help with that, lmk!
This makes no sense and impacts a lot of early stage software projects.
In many cases for bootstrap ones it makes them not financially viable - 100%+ effective tax rate would do that if you optimize for early profitability.
Even for VC backed ones, those taxes just eat into your runway. I know a bunch of companies that opt for some weird international setup to try and avoid the effects but it’s really not what you should do in early days…
Yeah, at this point a bunch of European countries have this setup. As long as you don’t have a complicated source-of-revenue situation (which is true for vast majority of population), you just confirm the tax return on a governmental app and you’re done. It’s extremely convenient, after they implemented a couple of years back, the time needed for tax return each year got down to maybe 2 min if you can use a computer.
Hashi got really good at ignoring PRs if they weren't their own. They even ignored the PRs coming from the dev teams of their own customers (ie users of TF Cloud and Enterprise) which speaks volumes about their willingness to listen to the community...
afaik you can't use terraform in the name, it's trademarked by Hashicorp
I've seen people using archbee.io for similar use case. Archbee is a lightweight docs tool for developers with various integrations so should check your boxes (full disclosure - I'm an investor there :)
That's an impressive time series and some pretty interesting results. How do you log that? How do you measure productivity?