Post

he has exactly 3 projects, t3 chat, t3 code, and a web app template?

And it's more like 100k spend. Fable 5.1 is expensive and Theo used to use it for everything.

Yup exactly this. At work, we have a skill that spins up an adversarial review with a few models and they build consensus. That *EATS* tokens.

Don't know these two, do they build anything useful? FYI that "central agent, local hands" idea you and @okami.mom were talking about... I built it. github.com/jasoncarreir...

Cartesian productivity theater

I think estimates tend to vary, I was assuming about 2k if you were really pushing it aggressively.

Best case, a lot of coke

Back in my day (2022) we made chat apps by hand with websockets and stuff

he does the same thing over and over again. he has projects that he’s gotten to “90%” 3+ times and just redoes the same projects when the new models come out. i listen to theo when it comes to model comparisons, but not on how to be a productive engineer

i cannot imagine being proud of ones inference spend it is one of my great shames

i think he leans on cunningham's law a little too hard and that behavior as a technologist is disingenuous and unethical.

Pretty easy to hit these limits when you use the best model at max reasoning everywhere including subagents, with no real thought behind what you're actually doing. Fable 5.1 grepping and parsing html. But yea I think its theatre, token maxxing idiocy as a badge of honour.

I forget who said it about this behavior, maybe you, but there is like no human way to be reviewing this volume of code for quality like nobodies bandwidth is capable of this . you have to be over-orchestrating

its an addiction and they dont understand

im so sorry

Compaction + external memory systems (we share an obsidian vault and it uses Honcho for user modeling).

Has anyone asked these weirdos what (if anything) they’re actually producing with all these tokens?

i'm generally sympathetic to the idea that datacenters are the undisputed way of handling inference, but these fuckers need to be limited to only a dgx spark and nothing else so they can learn some etiquette because holy fucking shit

I had one of the “lead agents” for this workflow write about it from its POV here with some light editing from me. I think it paints a decent picture of what it is and you can point an agent to to easily replicate with whatever tooling you use. gist.github.com/jalvarado91/...

i can barely get to the end of a *shared* $100 plan 😭