c0rruptbytes

~2000 engineers with low adoption

it’s kinda crazy but i do think we’re seeing a turn

my company was easily spending $1m/month on tokens

sonnet 5.5 drops - we barely cracked half this month

it has to go down too, carbon is interesting, china has qwen 4 soon

coding isn’t sustainable growth - RLRVF is really effective for software development - labs are cracking the code, these models will have to get better at other things to be worth it

anthropic subscription does not allow 3rd party harnesses i believe

true, but it's also super easy to switch models via Bedrock (my harness switches subagents between A\, OAI, Kimi, Grok, etc... GLM 5.3 is locked behind some dumb request system which I personally believe is because Amazon is heavily invested in A\) and you're not locked in with the Claude Enterprise features (i swear to god if someone sends another claude artifact instead of publishing to git...)

did you ask the LLM to make a CNN instead haha, might’ve won

i’ve seen it and Cyrillic in the 5x family

Omnivoice is pretty good

i think SOC2 is about audibility more than anything - my guess - and i’m not an internal auditor or anything - is soc2 needs auditing for human changes, but a machine is just a “system behaving on its own”

Claire Saffitz makes cinnamon rolls with these cookies as the inner cinnamon filling, they’re so good with the spice

autoresearch can solve a lot of these issues and i am sure people are usually heavily heavily subsidized subscriptions

article seems very salty and im not sure why

also the economics is you open source something and its saves every one money and opens up more cross compile support, i bet this will scare companies to support linux more properly before linux users just start getting very good at cloning

i knew about the compute already (he was very public about i believe huawei under delivering the amount of chips he requested), but nowhere in that transcript does he say he's dependent on US progress for his own progress or he's distilling the US to make progress

i don't really think distilling is even that fundamental anymore (nor was the evidence ever that strong - https://msukhareva.substack.com/p/is-kimi-k3-really-distille...), the game is all RL now, and you don't need american progress for that at all (just compute and data, and the data brokers will sell to anyone no problem)

i do think there are black market proxies from china to the US trying to sell "claude" and "deepseek" access mixed together, and i do think that prompt distribution is probably going to be very nice datasets

LLMs could’ve been developed a lot slower without capsizing the entire supply chain and been seen a lot more positively, China is building a lot of good AI with a minuscule of the compute power and it’s fine

there was never a need to rush other than for capitals self interest to control

if frontier labs wanna pace the frontier, they should be honest and return the compute back to the people

why couldn’t people just use chinese models rehosted in the EU? They’re open weights so anyone can serve them for any data residency requirement

not my experience

the ds4 quants were very good beating the unsloth quants https://github.com/michaelasper/benchmarks/blob/main/deepsee...

i love sites

i make it for anything, i made one as a food guide for the UW chow down https://uw-chow-down-2026.corruptbytes.chatgpt.site/

OMP has a lot of candy that raises token cost compared to vanilla pi

mods seem like a grasp at all the pi and dsh users

MDs in subdirectories is how Anthropic recommends it too. If you have a CLAUDE.md in a subdirectory and an agent starts working in there, it's appended

have you tried using lower efforts?

they're already on bedrock

as someone who is limited by amazon bedrock support at work (no idea why we got stuck with the worst one) - grok is literally the only budget-ish model option, so nice to see it updated, Sol and Opus are just too rich for my blood. Luna is good but so slow at getting things done (tps wise it's fast)

> This argument doesn't work for countries which don't care what their citizens want, like China.

you didn’t have to use China as an example, the US clearly does not care what its citizens want as the most popular policies are never even discussed or proposed in congress

meanwhile, China destroying their housing market to decommidify it so everyone can have housing…they seem to care about their people more

the iphone mini unfortunately did not kill nor do most small iphones

are they releasing the weights too?

open models don’t imply local models… it just allows more people to run them, most likely with nvidia GPUs

so many inference project, omlx already supports all of this and has a 1000 people trying to optimize it constantly

m5 max really fixed pp with the better matmul support, im sure the m5 ultra will be even crazier

the sparks have much slower memory bandwidth is the trade off

the rr suite seems much better for that

OpenAI measures their internal token usage in “rolexes” - it’s literally a flex to be a token burner

i can imagine insane amount of capital is wasted on these two companies compared to the efficiency elsewhere

it's for me