how i scaled my software output to 1,000 PRs/day and 260B tokens/month: -i was doing 50 PRs/day at the end of august. just a handful of Codex subscriptions, one laptop running 24/7 -took me 2 weeks to build reusable infra and autonomous verification systems. i decided upfront that fancy agent orchestration systems would get in the way of prompting more directly. i like to steer my agents in realtime so i use a lot of parallel terminal panes. -then i started layering on projects. i've accumulated a backlog of related ideas to explore. i started by picking up the simplest ones. i was adding 1-2 projects a day in the beginning (i have 81 repos now) -i started running out of subscription usage limits faster. first i was rotating every couple days. then it was every day. now it's multiple times a day. i peaked at 32 AI subscriptions across Codex, Claude, Devin, and Cursor. i'm not worried about this - initial buildout of software projects always takes the most code. after 1 month and 260B tokens, i ended up with 26 projects. several of them are pretty cool and ambitious, and all of them are useful (i use a lot of them daily). check out hraness. com -i find myself gravitating toward projects with lofty but verifiable goals: achieving SOTA on a benchmark, finding novel insights in data, optimizing token usage. it's easy to let agents cook on R&D if you're a good director. -it's not actually that hard to run 32 agents in parallel if all the threads are long-running (hours or days). few seem to understand this based on the reactions to videos i post of my setup. surprisingly even AI-pilled folks seem to take issue with tokenmaxxing at my scale. the sentiment is like: "i like coding with agents too but your race car is an abomination". i'm willing to bet that in a year or two, every AI builder will catch up to me and be spending 250B tokens/month (at a much more affordable price).

new personal record; five chatgpt 20x subscriptions with resets smoked and exhausted in 24 hours. thus it was ten 20x subs consumed 💀

left: me right: the haters

tokenmaxxing in my software factory like theres no tomorrow because its literally the end times for our profession and nobody has a fucking clue whats coming so im just building like a madman trying to find leverage and stay ahead of the curve glhf yall 32x Opus 5.5 today btw

ceo dropped into the architecture review said we should be building living applications auditable, self improving, programs that grow someone asked what that means in practice he said i'm giving you the vision, you give me the how so far the how is a cron job with a good attitude

today in my software factory i'm rocking Astra, Opus 5.5, and a ton of free SWE-2 on @DevinAI one giant @zeddotdev window with 19 agents adding TLA+ and Lean in a bunch of places i have 2 other monitors like this btw

Ding, ding, ding. AI is a time compression device that allows you to try many different ideas. Doesn't mean those ideas ship, but it means that you can do unhinged things to explore the idea space of possibilities. Sometimes just knowing something is possible is a valuable outcome. Being able to then compose that knowledge, that unique insight together, is an unhinged power. Knowing that A plus B compose well means that you can actually compose that to do C, and C actually ships.

btw while my agents were cooking i played tennis, watched the sun set and now i'm going to salsa class

my software factory uses 5-10B tokens a day 24x @DevinAI SWE-2 Max 6x Codex Astra Ultra 2x Claude Code Fable Max running in @zeddotdev

"we must decelerate now! we can't keep up with you and we're scared" mwahahaha i haven't even started, i'm just warming up

my mom lol

We must face the fact that we are on the brink of times when man may be able to magnify his intellectual and inventive capability, just as in the nineteenth century he used machines to magnify his physical capacity! Again, as then, our innocence is lost. And again, of course, the innocence, once lost, cannot be regained. The loss demands attention, not denial. – Christopher Alexander, Notes on the Synthesis of Form (1964)

$200 ChatGPT Pro 20x plan vs $100 5X plan

i just used up 15 Pro 20x subs over the weekend and i really needed to see this guess i'll buy more @DevinAI Max subs instead

every AI company right now is like wait Jev went viral for using structured output on a small model we were doing this in 2023 we're so back baby

late night software factory vibes 15x @DevinAI SWE-2 (free!) panelmaxxing in my beloved terminal @zeddotdev tonight we're building an agentic programming language/VM named ALGAL 🦠

read the latest dispatch from my software factory. banner generated with slopcamera.com