token math, and why i don't want a normal swe job

everything i've written here so far has been a build log: here's the thing, here's what shipped. this one's different. it's just what's actually on my mind this week, written down instead of scattered across a handful of x posts and forgotten.

token usage, differently

sonnet 5 just got launched. i live inside claude code for most of my day, the day job, byk, this site, so i've been testing it out today with real work. it's not what got me thinking, though. what it triggered is more interesting than the release itself.

i've started thinking about tokens the way i think about compute cost anywhere else: not "is this expensive" but "what's the cheapest way to test this idea." an experiment that used to cost me a day now costs some tokens and ten minutes of supervision. the real bottleneck moved from "can i build this" to "do i actually want to build this," which is a much better problem to have.

the failure mode i'm watching for in myself is throwing agents at a problem i haven't thought through yet, purely because it's cheap to try. cheap iteration is a gift until it quietly replaces thinking. i want the judgment call, whether something is worth doing, to stay mine. the "now go do it" part is what i hand off.

building this out loud

i only started posting anything publicly this week. moveworth was the first real thing i've ever put out, and it's still the only post i have.

it's been more energizing than i expected. watching other people build in public, posting the messy version without waiting for a "real" launch or anyone's permission, is what's pulling me to do more of it. so this is that: writing down what's actually on my mind instead of waiting until it feels finished enough to justify a post.

putting a take somewhere strangers can disagree with it is a different kind of uncomfortable than shipping a product, though. a product either works or it doesn't. an opinion just sits there and lets people judge you. doing it anyway, because distribution doesn't compound if it starts a year from now instead of today.

the market, briefly

everyone has a strong opinion right now about whether this is a bubble or the biggest platform shift since the internet showed up. i think both things are probably true at the same time. a lot of the money flowing in is chasing a story, and the underlying capability is still moving fast enough that last year's impressive demo is this year's thing i use before lunch. the tell isn't the hype. it's whether the tools keep getting better once the hype cycle wanders off to the next thing.

why i don't want a normal swe team

i'm effectively the only technical person inside a legal/compliance function at a major european tech company. i own problems end to end: i find them, scope them, build them, then watch someone actually use the result the next day.

on paper that sounds worse than a normal engineering seat. there's no queue of senior engineers reviewing my work, so i'm finding my own blind spots as i go. in practice it's taught me more in a few months than a standard graduate program would have, because the bar isn't "does this pass review," it's "does this survive contact with someone who has a real, messy, non-technical problem and doesn't care how clever the code is." legal and compliance work is full of exactly that kind of problem, and almost nobody is building for it. that's the actual edge. the sector is just where i happened to find it. being the only builder in a room full of non-builders forces a different kind of engineering than being the fortieth engineer on a team that already has strong opinions about everything.

the part i didn't expect: this is basically training for running my own company one day. finding a real problem nobody assigned me, deciding it's actually worth solving, building it, and owning what happens after someone starts depending on it, that's a founder's rep, not an engineer's rep. a normal swe team teaches you to build well inside someone else's decisions. this teaches you to make the decisions and live with them.

i'd want a team eventually, mostly for the mentorship. right now, not yet.

that's the notebook dump. more build logs coming. if you disagree with any of this, tell me: @benbykovski.