--- Log opened Wed Dec 07 00:00:38 2022
00:56 -!- human_g33k [~HumanG33k@2a01:e0a:95:5d90:215:c5ff:fe68:fb04] has joined #hplusroadmap
03:43 < pasky> kanzure: I'm not sure! I think most of the contributors now are the generic "core gnu contributors" that work on gcc, binutils .. many of them simply work for redhat, there are some working for intel/amd/arm working on optimization
03:43 < pasky> & new hardware support
05:24 < kanzure> alright.
05:45 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0::a324] has joined #hplusroadmap
05:57 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has quit [Ping timeout: 252 seconds]
06:11 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
06:31 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has joined #hplusroadmap
09:53 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has quit [Quit: oornbuten uustaple  ould qe oosmic nay]
10:10 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
11:41 -!- Malvolio is now known as ANACHRON
12:51 -!- ANACHRON [~Malvolio@idlerpg/player/Malvolio] has quit [Remote host closed the connection]
13:57 -!- codaraxis [~codaraxis@user/codaraxis] has joined #hplusroadmap
14:22 < kanzure> hmph
14:27 < kanzure> if chatgpt has a 45 terabyte common crawl data set that it is trained off of, how many terabytes of chatgpt interaction would be required to train another gpt-3 model from that? in the order of the number of queries to chatgpt
14:29 < kanzure> according to GPT-3 playground, it's telling me this would need only 10 terabytes of chatgpt session logs
14:30 < kanzure> which is probably wrong, but maybe there is a way to sample the chatgpt model such that you treat it as a compression function over the original training data set (which probably had lots of redundancies)
15:02 -!- srk- [~sorki@user/srk] has joined #hplusroadmap
15:05 -!- srk [~sorki@user/srk] has quit [Ping timeout: 260 seconds]
15:05 -!- srk- is now known as srk
15:22 -!- srk- [~sorki@user/srk] has joined #hplusroadmap
15:22 -!- srk [~sorki@user/srk] has quit [Ping timeout: 268 seconds]
15:24 -!- srk- is now known as srk
16:07 < fenn> there's the problem of coming up with prompts to sample the entire latent space
16:08 < fenn> it's not necessarily all connected, so you can't just spider like a webcrawler
16:10 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has quit [Ping timeout: 268 seconds]
16:16 < pasky> the prompt is a vector, it might be good enough to just sample a grid in the vector space of prompts and it's a question then of how fine a grid you need
16:16 < pasky> assuming the internal prompt vector has the same size as embeddings (just a guess), that'd be 12288-vector, hm. never mind i guess :)
16:17 < pasky> this actually wouldn't work with any reasonably sized vector, you need to compress the vector space
16:24 < kanzure> is it plausible to steal their parameters this way?
16:34 < pasky> well your actual gpt input is the prompt, not the encoded vector; this would require direct access to the model
16:38 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has left #hplusroadmap []
16:57 < TMA> pasky: do not look backwards. good condition handling is already present in common lisp, therefore obsolete :) (like if with an else branch, anonymous functions (lambda), garbage collector, ...)
17:02 < pasky> TMA: er what are you replying to? :)
17:02 < TMA> pasky: your tweet series, of course https://twitter.com/xpasky/status/1596618565167513601
17:17 -!- human_g33k [~HumanG33k@2a01:e0a:95:5d90:215:c5ff:fe68:fb04] has quit [Ping timeout: 260 seconds]
17:18 -!- human_g33k [~HumanG33k@2a01:e0a:95:5d90:215:c5ff:fe68:fb04] has joined #hplusroadmap
18:30 < kanzure> "Precise atom manipulation through deep reinforcement learning" https://www.nature.com/articles/s41467-022-35149-w
18:34 -!- codaraxis__ [~codaraxis@user/codaraxis] has joined #hplusroadmap
18:34 < kanzure> "A plant-derived natural photosynthetic system for improving cell anabolism" https://www.nature.com/articles/s41586-022-05499-y
18:36 -!- codaraxis [~codaraxis@user/codaraxis] has quit [Ping timeout: 260 seconds]
18:38 -!- codaraxis__ [~codaraxis@user/codaraxis] has quit [Ping timeout: 256 seconds]
18:41 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
19:05 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has quit [Killed (zirconium.libera.chat (Nickname regained by services))]
19:06 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
19:17 -!- test__ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
19:21 -!- test_ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 260 seconds]
19:23 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0::a324] has quit [Quit: Leaving]
19:25 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has quit [Quit: everyone loves the new machine]
19:43 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
20:14 -!- ANACHRON [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
21:03 < lsneff> well, it’s official, i’m coming back to spacex after i graduate
21:06 -!- ANACHRON [~Malvolio@idlerpg/player/Malvolio] has quit [Read error: Connection reset by peer]
21:06 -!- srk [~sorki@user/srk] has quit [Ping timeout: 268 seconds]
21:09 -!- ANACHRON [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
21:18 -!- ANACHRON [~Malvolio@idlerpg/player/Malvolio] has quit [Read error: Connection reset by peer]
21:42 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has quit [Ping timeout: 260 seconds]
21:57 < fenn> gratz. i'll keep an eye out for you on the labpadre livestream :)
21:58 < fenn> that speck on the manlift in the rain at 4 am ahahahaha
22:09 < lsneff> :)
22:12 < lsneff> my coworkers were joking it’d be more beneficial for direct-to-cell if I went to starbase to work on starship instead of working on stuff up in seattle lol
22:23 -!- codaraxis [~codaraxis@user/codaraxis] has joined #hplusroadmap
23:59 -!- superkuh [~superkuh@user/superkuh] has quit [Ping timeout: 252 seconds]
--- Log closed Thu Dec 08 00:00:39 2022