Nick

@kukkurovaca

he/they villainx
flickr.com/photos/kukkurovaca

objectively too many tv/anime/manga screencaps

photography, especially infrared

weird keyboards

new music listening notes

no nazis, no terfs, no yimbies

Flickr

www.flickr.com/photos/kukkurovaca/

Alt: The Cohost Daily Tag

cohost.org/TagOfTheDay

Photoblog

www.nickshere.com/

Landing Page

kukkurovaca.com/

Obsidian Vault / Psuedoblog

plaintextadventure.com/

Combined RSS Feed

kukkurovaca.com/rss.xml

OPML

kukkurovaca.com/feeds.opml

Nick @kukkurovaca2/22/2023, 7:17 PM

colin @vogon

colin @vogon2/22/2023, 6:34 PM

man I kind of understood when everyone was adding NFTs to their products for no reason except that it sounded trendy, because at the very least it's not gonna make you any money but it's also not really gonna cost you any money

the shit people are pulling with large language models over the past couple months? no fucking clue, unless OpenAI is rolling dump trucks of money and service credits up to everyone's door in the sincere but dreadfully misguided thought that this will become ubiquitous in a couple years, at which point they'll all be locked into contracts

colin @vogon2/22/2023, 6:51 PM

I posted this before I found out that Spotify launched a "DJ" today that does generative banter so just imagine that I posted this again, underlined

You must log in to comment.

in reply to @vogon's post:

Big Mick Energy @possumskull2/22/2023, 6:39 PM

This is one thing I haven't seen a lot of talk about, the fact that large language models are a bear to train and use, just absolutely enormous computational expense

ash @hierarchon2/22/2023, 8:13 PM

it's really funny that image models can be feasibly run on a consumer GPU but text models are completely out of reach. you'd think it'd be the opposite

stu @stu2/22/2023, 8:51 PM

a picture truly is worth a thousand words

roy @corhocysen2/22/2023, 6:58 PM

i thought the mining fees for NFTs were pretty significant? i suppose not compared to sale prices at the height of the hype, but still

colin @vogon2/22/2023, 7:00 PM

they were! but people quickly found a way around having to pay for the mining fees up front, and only once a sucker^W valued customer decided to buy

Tenna Lotor @tenna2/22/2023, 7:01 PM

Depends on how you're doing it. If you're using one of the smaller chains or a "sidechain" I've seen a few companies use, it's pennies if even that.

Alternatively, as colin said, it can be one of those things where the mint price is paid upon purchase of it, so it's something the selling company doesn't have to pay for at all - just the consumer.

Nire Bryce @NireBryce2/22/2023, 7:22 PM

Microsoft is pouring money into it, and it runs on azure so they're probably also undercutting the price, I assume because they are trying to force Google to release LaMDA while it's still half baked

Gravity Pike @gravity-pike2/22/2023, 9:11 PM

Microsoft dumped $10B into OpenAI and got Sydney: The Chatbot Who Can Gaslight You, at the same time that they fired thousands of real employees.

Capital is out here fucking DEMANDING that you gamble big on magic beans, because there's no theoretical limit to what kind of magic could be in 'em.

gender reveal maenad @phorusrhacid2/22/2023, 9:39 PM

i am so excited for cvs' automated voice thing that spams advertisements at you while trying to not let you talk to a pharmacist for ten seconds to get even worse at processing requests

in reply to @vogon's post:

2 mello @MelloMakes2/22/2023, 6:57 PM

excuse me?!?! lol, they should bring back radio bumpers too. everyone who switched to streaming really misses things interrupting music

colin @vogon2/22/2023, 6:58 PM

yeah like the history of music radio from the '80s onward was that every station ident mentions how little talking they're going to be doing over/in between the music and they know everyone hates The Other Stations that do a lot of it

unless you've never turned on a radio in your life, I can't imagine someone deciding to pay money to intentionally add in banter

@gravely2/22/2023, 6:59 PM

from the outside, these calls seem like the decisions of fewer and more cynical leaders higher in org charts who are uninformed and blindly trend-following and/or uninformed and failing to say no to bad ideas (which is imho their actual primary job all day).

Molly @Yerfriendmolly2/22/2023, 7:03 PM

i think one of the funniest things is people on various boards posting fake leaks from upcoming streams like 'the next pokemon game will use open ai technology to generate new dungeons' and like dogg. no it won't. you're out of your fucking mind

Tenna Lotor @tenna2/22/2023, 7:05 PM

It really does make you wonder

ur bff jll @wolfwood2/22/2023, 7:28 PM

omg whyyy. but also maybe we need a concerted effort to highlight the energy consumed by training, as mentioned in another comment.

everyone posting about b*tco*n burning down the rainforest, but with AI.

also I'm never gonna get to buy a new gpu ever again am I?

one has to wonder with Intel joining the space and AMD being increasingly relevant, who is striking massive accelerator deals to fuel all this? has CUDA lock in kept it all on nvidia? they are certainly feeling a lot of investor pressure to find a market to replace tulip mania hypergrowth.

how much are silicon vendor wagging this dog?

colin @vogon2/22/2023, 7:33 PM

so, the one bit of good news is that at this point, afaik most of the serious ML platform-as-a-service providers have abandoned GPUs as a computational substrate in favor of dedicated tensor accelerators, because they consume about 1/10 as much power per operation and throw off proportionally less waste heat, so they're much cheaper and easier to feed -- so at least you're not in competition with amazon and google for GPUs specifically any more, just glue circuitry and fab capacity.

still sucks though!

ur bff jll @wolfwood2/22/2023, 8:18 PM

still competing for high speed memory? which I suspect is a becoming bigger chunk of a 16 GB card's price than the tiering indicates.

and yeah, fab capacity :( semiconducting 3D printer filament when?

colin @vogon2/22/2023, 8:23 PM

ah yeah! good point, I hadn't considered RAM. most of the real tensor accelerators are non-public hardware so it's difficult to speculate, though google does sell a consumerized one in the form of Coral (https://www.coral.ai/products/dev-board), which only uses LPDDR4 right now rather than (G)DDR5.

ur bff jll @wolfwood2/22/2023, 8:31 PM

hmm, the 1 or 4 GB LPDDR4 is for the NXP chip. is the tensor doodad piggybacking on the system memory over PCIe? I'd have expected it to have significant local memory, but maybe it's only using SRAM caches because data has good access properties?

colin @vogon2/22/2023, 8:34 PM

oh yeah, seems so from the datasheet:

Uses PCIe and I2C/GPIO to interface with the iMX 8M SoC

https://www.coral.ai/static/files/Coral-Dev-Board-datasheet.pdf

Star's Curiosity Bazaar @Star-Feline2/22/2023, 8:16 PM

Alright I'm gonna say a DJ feature would be pretty good if it's toggleable and the AI is only focused on giving the history of the music while you're listening instead of idle chit chat and the weather.

The thing is AI has a lot of actual tangible uses, which is something that No Fucking Thanks can't say for themselves. I am quite worried about the energy drainage though, I'm not sure how sustainable this AI gold rush is.

signal eleven @widr2/22/2023, 8:32 PM

i have a really sick bad feeling that there is a lot of genuine interest and demand in the near term for this shit. already i'm unable to drive home with people why LLM code assistance is a bad idea. "what about when it gets 3x better? 10x better? what about when it's 1000x better than it is now???" i'm so tired

somewhat nifty @somewhatnifty2/22/2023, 9:23 PM

looking forward to this going the same place as infinite seinfeld

ABC @ABC2/22/2023, 9:28 PM

I'd actually use it if the Spotify algorithms were halfway decent. Every time I boot Spotify they're like "Ahh we see you like rap, how about 5 Macklemore tracks in a row"

Pinned Tags