--- Log opened Sat Jan 07 00:00:07 2023
00:51 -!- stipa_ [~stipa@user/stipa] has joined #hplusroadmap
00:53 -!- stipa [~stipa@user/stipa] has quit [Ping timeout: 256 seconds]
00:53 -!- stipa_ is now known as stipa
01:44 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
01:45 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
04:40 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0::4249] has joined #hplusroadmap
06:36 -!- cthlolo [~lorogue@77.33.23.154.dhcp.fibianet.dk] has joined #hplusroadmap
08:07 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
08:07 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
12:29 < nsh> https://www.youtube.com/watch?v=pJT1CKuS378
12:29 < Muaddib> [pJT1CKuS378] Seeds of a Good Anthropocene | Elena Bennett (79:11)
12:30 < nsh> (strongly recommended)
13:18 < docl> it occurs to me that if ML models are able to become convincingly sapient (doesn't have to be real), people might adopt models as children. makes at least as much sense as adopting a small dog
13:24 < kanzure> docl: might be too expensive in computational resources/power for people to really adopt. humans are quite cheap in comparison.
13:26 < docl> I wonder what some reasonable comparisons are by the numbers? Supposedly chatGPT convos only cost a few cents each
13:26 < docl> but having a separate model per kid might drive up the cost overall?
13:34 < kanzure> http://www.forth.org/svfig/
13:34 < kanzure> docl: that's a few cents per query not for training
13:37 < docl> hmm. I foresee there being a lot of caveats to this question. e.g. a starter model that's pretrained to be trainable along humanlike lines without a huge daily expenditure
13:38 < docl> $10/day in electricity say
13:38 < kanzure> i think this was one of my criticisms of robbie hanson's emulations book
13:38 < kanzure> we don't know how much this stuff will cost to run
13:39 < kanzure> and the cost might be rather high for a single emulation of a single brain
13:39 < docl> yeah but brain emulations probably have a lot of legacy biology stuff not needed for convincingly-sapient AI models
13:41 < docl> I think there's probably several orders magnitude greater expense for something based on emulating the function of literal neurons
13:41 < kanzure> i was trying to find some cognitive architecture software stuff that isn't based on neuron emulation or spiking neuron models
13:41 < kanzure> everything is too derivative of the artificial neuron network concept
13:42 < fenn> each task only needs to be learned by one AI
13:42 < docl> (I'm just guessing on general principles, could be wrong)
13:43 < kanzure> hippocampus apparently can be modeled as a series of differential equations
13:43 < kanzure> and im not sure what the auditory cortex is doing but i can't imagine it absolutely requires a neural network
13:43 < fenn> remember furby? the creepy talking furry monster things. they faked learning by gradually improving the performance of the software
13:44 < docl> I never actually met a furby, heard about them but wasn't aware they fake learned
13:45 < kanzure> docl: do you think 3d printing of human brain tissue would be faster than developing software artificial general intelligence
13:45 < fenn> they have a pre-set vocabulary that increases as you interact with them
13:53 < docl> kanzure: it seems like we're falling behind on the keep-brain-tissue-alive-in-a-jar vs the improve-ml-to-generalize-more race
13:54 < docl> so probably not? but maybe, contingent on e.g. something brains generalize well on not being at all easy for software
13:55 < kanzure> is there a problem with neural tissue cultures longevity?
13:57 < docl> I haven't heard of a whole brain being kept alive for more than a few hours. but maybe that's just people being shy about publicity
13:57 < kanzure> you may not need a whole brain
13:57 < docl> maybe just printing a mess of neurons in 2d sheets on some electronics and networking them is a better approack
13:58 < kanzure> yeah
13:58 < kanzure> but (1) people have done that, and (2) there have been companies, and (3) none of the results have sounded impressive yet and i don't now why
13:58 < kanzure> *i don't know why
13:58 < docl> thin sheets would let you replace the CSF without having to deal with something like a vascular system
13:59 < kanzure> like the papers in the 90s were "we got these neurons in a petri dish to pilot a flight simulator" and it turns out it's like 8 inputs and 3 outputs which is not interesting
14:00 < kanzure> (and also, i'm not sure but once you have a plane flying in a flight simulator, isn't it easy to keep it in the air?)
14:00 < docl> huh. you'd think the first thing people would try is make a few square meters of electronics+neurons and try to teach it to talk
14:01 < fenn> how do you reward/punish a sheet of neurons?
14:02 < fenn> seems like it would just predict what happens next, since that's all the cortex really does
14:03 < kanzure> i don't know how labs do that, but, you could do thermoregulation where you try to drive the neurons to minimize the heat you are adding into the system (eg out of homeostasis for one of the input cells)
14:14 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has left #hplusroadmap []
14:15 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
14:16 < kanzure> wonder if you could pattern a microfluidic chip with channels that only axons or dendrites would grow into
14:17 < kanzure> (you would print embryonic neuronal stem cells at each location before you close the device)
14:17 < kanzure> surely someone did something beyond flight control systems with their in vitro neural tissue cultures
14:38 < docl> I wonder if something like planarian neurons would be easier to keep alive in vitro than human neurons
14:41 < kanzure> why did the cortical labs dishbrain stop at pong? https://www.nature.com/articles/d41586-022-03229-y https://www.sciencedirect.com/science/article/pii/S0896627322008066 https://corticallabs.com/
14:42 < kanzure> took them 3 years apparently https://corticallabs.medium.com/in-vitro-neurons-learn-and-exhibit-sentience-when-embodied-in-a-simulated-gameworld-387ec3f2c870
14:42 < docl> also wonder whether there is any difference in responsiveness to training based on what species you get the neuron from. could parrot neurons in a system like this be taught to talk more easily than hamster neurons? (my guess is no, they are not learning in an analogous enough way to nature for that to matter)
14:45 < kanzure> dunno why this is all using microelectrodes- maybe there's no rejection, but optogenetics + chromatophores should work fine
15:22 -!- cthlolo [~lorogue@77.33.23.154.dhcp.fibianet.dk] has quit [Read error: Connection reset by peer]
16:02 < kanzure> altho they do report being able to learn pong in ~5 minutes
16:57 < kanzure> minicircle is still doing things https://minicircle.clinic/our-therapies/
17:07 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has quit [Read error: Connection reset by peer]
17:47 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
18:14 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has quit [Read error: Connection reset by peer]
18:25 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0::4249] has quit [Quit: Leaving]
19:00 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
19:02 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has quit [Read error: Connection reset by peer]
19:08 -!- Jay_Dugger [~jwd@47.189.1.51] has quit [Quit: Leaving]
20:29 < lsneff> looks like im gonna finish my undergrad thesis over the summer
20:30 < lsneff> really like my major, but really wish I could do a normal capstone like most engineering atudents
20:44 < muurkha> that's great
20:44 < muurkha> why can't you?
22:25 -!- mlaga97 [~quassel@user/mlaga97] has quit [Ping timeout: 272 seconds]
22:41 -!- mlaga97 [~quassel@user/mlaga97] has joined #hplusroadmap
--- Log closed Sun Jan 08 00:00:08 2023