--- Log opened Sun Apr 20 00:00:21 2025 00:38 -!- TMA [tma@twin.jikos.cz] has quit [Ping timeout: 252 seconds] 00:45 < hprmbridge> jrayhawk> pinging @mark for SF/Austin person with an explicitly stated interest in giving away compute hours to molecular simulation experiments https://x.com/ctjlewis/status/1913698680491975112 01:02 -!- Jenda [~jenda@coralmyn.hrach.eu] has quit [Ping timeout: 260 seconds] 01:10 -!- Jenda [~jenda@coralmyn.hrach.eu] has joined #hplusroadmap 01:14 -!- Jenda [~jenda@coralmyn.hrach.eu] has quit [Client Quit] 01:14 -!- Jenda [~jenda@coralmyn.hrach.eu] has joined #hplusroadmap 01:16 -!- Jenda [~jenda@coralmyn.hrach.eu] has quit [Client Quit] 01:16 -!- Jenda [~jenda@coralmyn.hrach.eu] has joined #hplusroadmap 01:21 < fenn> seeing plots like this just makes me think "they said it couldn't be done..." https://diffuse.one/images/hypothetical_limits.png 01:23 < fenn> (the graph is limits on AI training compute before the earth melts so not counting extra-terrestrial energy sources like dyson spheres) 01:25 < fenn> from https://diffuse.one/p/d1-001 01:25 < nsh> 'tain't what ye do it's the way that ye dae it 01:27 < fenn> "not some exotic device like a ... cryogenic computer" uhh 01:28 < fenn> doesn't seem all that exotic 01:30 < fenn> i was arguing with jesse about limits of heat rejection when you scale up energy production. yeah if you use black solar panels there's a limit on passive heat rejection, but nuclear reactors can radiate directly to space and not use the atmosphere as a heat sink. 01:30 < fenn> even solar panels can pump heat around in exchange for an efficiency hit 01:31 < fenn> it's not like we're running steam engines on the solar panels' thermal output 01:33 < fenn> then there's the end use side. if all we're doing is running GPUs then they can also radiate directly into space by insulating the radiator from the ground 01:34 < fenn> it costs more which is why nobody does this today, also there is no reason to do so at the tiny global energy use of 2025 01:37 < fenn> i sorta wonder if AI was solved in the 1980s and it just required special purpose hardware and they didn't want to let anyone else use the tech and are all laughing at us doing it wrong but can't stop it because the wrong way is inevitable 01:38 < fenn> you can binarize neural networks to run directly in digital logic extremely fast and efficiently, and can probably train them directly by averaging large batch sizes instead of using gradient descent 01:40 < fenn> like maybe that icegods guy was not a crank 01:40 < fenn> related: 01:40 < fenn> .t https://www.youtube.com/watch?v=FKQfMwFZvIE 01:40 < fenn> saxo: my dude 01:41 < fenn> [ Convolutional Differentiable Logic Gate Networks - NeurIPS Oral - difflogic (5m1s) Felix Petersen ] 01:42 < fenn> .m https://x.com/_brickner/status/1871348156786704657 01:42 < AugustaAva> ​twitter: <_brickner> wrote a paper: it lets you *train* in 1.58b! could use 97% less energy, 90% less weight memory. leads to a new model format which can store a 175B model in ~20mb. also, no backprop! https://pbs.twimg.com/media/GfhTeN7XcAAN1Pf.png 01:43 < fenn> "1.58b" is weight values in {-1,0,+1} 01:45 < fenn> "NoiseStep" because you're relying on the randomness inherent in the training data to smooth out binary activations, i guess 01:50 < fenn> he also adds noise 01:51 < fenn> we don't actually need the extra .58b and can just use {-1,+1} which sum to zero when added together, for a small increase in weight count 01:53 < fenn> this makes it easy to directly implement in transistor logic which is well understood, fast, inexpensive 01:53 < fenn> i forget what my point was. anyway, FPGAs are cool, let's do more of those 02:59 < nsh> .m https://x.com/dbenner83/status/1913585746541228217 02:59 < AugustaAva> ​twitter: Today in 1993, after a senseless siege that lasted 51 days, the US government murdered 82 people at Waco. ␤ ␤ US Attorney General Janet Reno authorized tanks to flood the compound with tear gas. ␤ ␤ Instead of encouraging evacuation, the decision backfired. ␤ ␤ Those inside scrambled to https://pbs.twimg.com/media/Go5rtfuXgAA0K_g.jpg 02:59 < nsh> nevar fargut 06:31 -!- stipa [~stipa@user/stipa] has quit [Quit: WeeChat 4.3.5] 06:47 -!- stipa [~stipa@user/stipa] has joined #hplusroadmap 06:51 -!- stipa [~stipa@user/stipa] has quit [Ping timeout: 252 seconds] 06:53 -!- stipa [~stipa@user/stipa] has joined #hplusroadmap 08:05 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.] 08:05 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap 08:30 < alethkit> FelixWeis__: Who is icegods? 11:37 < hprmbridge> lordkek__> https://docs.google.com/document/d/19lP57SK2RXbdvkbC1L_U-hRldEkwSNjEcCdmtN3tIMo/edit 11:47 < hprmbridge> kanzure> no thanks 12:31 < NewtonTrendy> i know someone with a phd who thinks llm text models are good enough save formats for the state of a human brain 12:55 < L29Ah> you americans have all sorts of funny phds 13:39 < hprmbridge> kanzure> why use an LLM to save state instead of writing the state to disk directly? 14:28 -!- nefercheprure [tma@twin.jikos.cz] has joined #hplusroadmap 14:43 < fenn> .m https://x.com/basedmachina/status/1902338020646224151 14:43 < AugustaAva> ​twitter: Most interesting effect of taking creatine is that I wake up READY, zero startup time, just go. 14:47 -!- nefercheprure is now known as TMA 14:51 < L29Ah> can you reproduce? 14:51 < L29Ah> er 14:51 < L29Ah> do you observe the same effect? 14:53 < fenn> L29Ah: only one day so far but yes. it could be placebo 14:54 < fenn> alethkit: https://web.archive.org/web/20110222185149/http://icegods.com/ 14:55 < fenn> L29Ah: immediately on waking i noticed a new feeling like a sort of internal hydrostatic pressure 14:56 < fenn> i still made a couple stupid errors like dropping my brush 14:57 < fenn> just 3g creatine before bed 15:32 < fenn> in the past i noticed a similar waking up ready feeling the day after taking modafinil --- Log closed Mon Apr 21 00:00:22 2025