I imagine some scenarios where you're throwing a big data set at an LLM could add up quickly.
"Check all of Confluence for outdated and conflicting info"
"Review all the legal contracts over the past 10 years"
"Evaluate the source code in all our dependencies for vulnerabilities"
All of those would be fairly straight forward for a single person to kick off without thinking about the data set size. Especially if they have Claude set to Opus 4.7 x high effort for everything.
Even someone saying something like "rewrite the 25 year old Java monolith in rust" as a PoC and leaving it running in the background for a week
But even that doesn't explain it. If you ask claude to read all of Confluence, it will read like 5-10 pages, read an overview of the workspace, grep a few things, then give you an unreliable summary.
You have to fight quite a lot with agents to get them to actually read the fucking files in full
In the corporate world, writing code was always a small chunk anyway. Iirc something like 30% of employee time. Getting 50% faster there still only gains back 10% of your time.
To further complicate matters, you had to spend on AI software and potentially additional on legal/risk/security/compliance to enable that.
The smaller the company, the bigger that % of coding is of total time (all the way down to hobby where the majority of time is spent coding)
Where I work, I would guess that, out of the amount of person-hours spent on a project, less than 5% of it is the cumulative time spent designing and writing code. There's so much other stuff going on: Requirements gathering, "aligning" with other teams, human reviews, QA, presenting to executives, waiting for approval from executives, reporting status, deploying to staging, internal dogfooding, slow ramp-ups to production... This is how projects that are 50 lines of code take 3 months to deploy. AI is helping reduce the time spent on that 2%.
Haven't used Vercel but back in the Heroku/CloudFoundry days it was pretty easy to jam arbitrary binaries into the runtime containers and some of the build packs were flexible enough you could override most of the functionality.
Not sure if that's possible/how easy it is on Vercel
I think they were doing that before AI got big the last couple years.
Their core network stuff always seemed pretty robust but all the newer stuff was much more thrown together. Thinking specifically of Zero Trust/Argo Tunnels which has been around a few years (and I do like) but has some rough edges.
Git is typically fairly slow if you have to wait on a test suite and deployment pipeline. Usually at least 10 minutes but sometimes 30, 60, 90+ minutes. A lot of purprose-built feature flag platforms hot reload the config in seconds.
JSON in the repo also risks introducing customer data to git if you want to rollout based on specific customer attributes (sometimes, for us, it's a list of early opt-in customers we have meetings with to discuss/develop new features)
It's also less accessible for "business users" like product/project managers, sales, and marketing they want to coordinate feature rollout with other business initiatives (and don't want to bother engineers when they do)
> "belt and suspenders"
reply