snow
@snowclipsed
I like surfing latent spaces.
just broke through the goal with a current gen speed of 16tok/s on an intel laptop CPU and upwards of 20+ tok/s on macs M1-M3 pro. your CPU has untapped potential :)
went from 8 to 11+ tok/s on a laptop CPU, onwards to 15 (thanks to a lot of PRs and more optimization work!) x.com/snowclipsed/st…
the impact that early slop LLMs have had on society had completely ruined some text(ing) patterns for me
> "entry level jobs" > "could he replace his lawyer" I'm sure being Dario Amodei's lawyer is not an entry level job
@DarioAmodei 's lawyer had to apologize for using a fake citation in her draft filings...generated by Claude. So while he says AI could eliminate HALF of all entry-level white-collar jobs within 5 years, would *he* replace his human lawyers with his billion-dollar LLM? (3/n)
ironically that infographic feels like it was written by AI? also, this take is terribly not future proof. a lot of the early automation is (naturally) moving towards lower-stakes, entry level jobs.
Hallucination is baked into LLMs. Can't be eliminated, it's how they work. @DarioAmodei says LLMs hallucinate less than humans. But it's not about less or more. It's the differing & dangerous nature of the hallucination, making it unlikely LLMs will cause mass unemployment (1/n)
who's making a bird neural network
Bird saves and reproduces data: A PNG image of a bird ( photo of a bird -> spectral synthesizer ) was reproduced by an adult Starling bird youtu.be/hCQCP-5g5bo?si…. It seems to have reproduced the sound in conjunction with some additional notes which made it not detectable…
hot take (?) : the way AI will take away all the 'low-tier software jobs' is by changing the new grad skill distribution and making them entirely reliant on it to code
26 months into 6 months away from AI stealing yo jerbs - End of April CEO of replit has announced that it's 6-18 months before no more engineers are needed - replit ai deleted someone's start up. It was saved by engineers
gpt-5
Lobster -> GPT-5 Nectarine -> GPT-5-mini Starfish -> GPT-5-nano
A feature for llm chat webuis that is very low-hanging fruit is using context isolation to start a new chat with all the relevant information you gathered in the current chat about a specific topic.
there's a really fun question that's been going around in my friend circles again : what percentage of X is bots?
"its not intellectually stimulating, so its not worth doing. i'm above working on this" is a trap that smart people seem to have a very, very easy time falling into
another crazy, completely unfiltered idea i had is - why not just condition the decoder to output multiple bytes? surely it's doable
ys u need to somehow have cheap communication between some stuff that is doing the smaller pass and some stuff that is doing the larger pass. or many concurrent larger passes