wireless
@wireless_anon
Followers
13K
Following
7K
Statuses
3K
stealth (50K DAU) | previously ML @ google, competitive programmer
Joined January 2022
@theo @thomas_chardo You can call Mistral Large 2.5 via their API without the license but then you wouldn't be able to self-host it on Cerebras which is necessary to achieve Le Chat's latency.
0
0
0
@_ashleypeacock @CloudflareDev We use separate CF accounts for dev / prod. It is pretty similar to the AWS org set up so far, but there is no unified billing (which also means that our enterprise features on our prod account aren't available in our dev account).
1
0
0
i think it normal webdev, there are too ways it could break. in react native isn't sort of ok since you don't have things like child/descendent/etc selectors. i once had a custom font (not supported by nativewind's builtin font handling), so i made a custom text component that detected font weight from classname / styles and applied it.
0
0
1
100b tokens is really small. keep in mind the average request is probably 2.5K+ input tokens given system prompt + user memory + multi-turn convos. reasoning models also produce a huge amount of output tokens, which consume more compute than input tokens. i'd also guess the "hundreds of billions" figure from cerebras is based on an optimistic scenario (e.g. 100b input tokens of llama8b per day) my rough guess based on internal data from a relatively popular chatgpt-like app is that 100b tokens/day is enough for ~800k DAU.
0
0
2
@thdxr i think the startups are in a weird spot. if the market expands enough, eventually openai / anthropic will take interest. they already have canvas feature, so directionally it's only a few iterations away.
1
0
1
@Simeon_Cps probably less profitable than making a model that tries to give an answer (even if it hallucinates some % of the time).
0
0
4
@catehall if you normally use the app, try checking the website. i didn't see it in the ios and mac apps and figured i didn't have access yet, then i checked the website and saw it there.
1
0
0
@hmzdot @RhysSullivan How do you replicate Effect-style dependency injection in Rust? Specifically the way it can infer, propagate, and enforce dependencies on the type level.
1
0
2
when i was at google (7 years ago), you would hear stories about people that independently "founded" a product within google and that went on to become super successful. gmail is a famous example. at the time, i got the sense that there was still some of that inventive spirit, but it had already declined significantly. now, 7 years later, i suspect that it is completely gone -- anyone who would be willing/capable of doing that is now a startup founder.
1
0
12
@teortaxesTex it seems like he implies implies sonnet was <$40M (he says "on trend" would be 8x cost reduction, and then says deepseek is behind trend). could be misinterpreting it though.
1
0
3
people are confusing the $5m figure with the the total amount spent on talent/experiments/data/etc. (at least ~1 oom higher) which kinda makes a lot of the discourse nonsensical.
0
0
6
@Teknium1 @xlr8harder how does it not get completely bottlenecked by interconnect on 256 gpus? are these hopper gpus?
0
0
0