Back

the reason why I don’t get imposter syndrome is because I will never spew stupid bullshit as loud or as rapid as the folks my industry thinks are knowledgeable

Currently we have single-threaded execution running at ~10Hz (tok/s) and enjoy looking at the assembly-level execution traces stream by.

starting to feel a lot better about the FPGA work I’ve been doing, if a 10Hz single-threaded processor is considered anywhere near a success

tfw your drunken friday night shellscripts are someone else’s whole startup model[0]

it is astounding just how convoluted and self-twisted these people manage to make their arguments though. like, this thing. literally billions of investment over time, multiple decamillions (if not more) in actual in-the-rack compute, however fucking much else you have in supporting infra (power, IP, routing, and all the other cumulative layer cake of goog’s investment in that over time), all as actual fucking computers that are probably already running virtual machines to run the containers for running your model code… and then you get this fucking kind of remark out of it

it takes some incredibly creative point-dodging

[0] - unironically I’ve had this feel a non-zero amount of of times

it’s fucking bizarre to see this many supposed computer science experts forget how much of the field is built on complexity theory and optimization, since things get fucking weird (simulate an entire alternate universe weird, like we’ve seen a lot of these AI fools turn to) when you stop analyzing complexity

of course my FPGA obsession lately has been designing a lambda calculus reducer that runs on hardware, a ridiculously inefficient (in terms of both cycles and bytes) way to do general computation. I don’t claim that work’s practical though, unlike these breathless AI fuckers pretending an LLM doing inefficient computation (on top of a multitude of layers of efficient computation) is revolutionary