--- Log opened Wed Jun 28 00:00:43 2023
00:29 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has joined #hplusroadmap
02:56 < hprmbridge> nmz787> Yeah you're gonna be dming with everyone in here it we weren't so against it
03:38 -!- EnabrinTain_ [sid11525@id-11525.helmsley.irccloud.com] has quit [Server closed connection]
03:38 -!- EnabrinTain_ [sid11525@id-11525.helmsley.irccloud.com] has joined #hplusroadmap
03:40 -!- test_ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
03:43 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 250 seconds]
04:13 -!- RubenSomsen [sid301948@user/rubensomsen] has quit [Server closed connection]
04:14 -!- RubenSomsen [sid301948@user/rubensomsen] has joined #hplusroadmap
04:20 -!- sgiath [~sgiath@mail.sgiath.dev] has quit [Ping timeout: 250 seconds]
04:20 -!- sgiath [~sgiath@mail.sgiath.dev] has joined #hplusroadmap
04:42 -!- test_ is now known as _flood
05:00 < hprmbridge> kanzure> kicad but with ai copilot https://twitter.com/BuildWithFlux/status/1673729156792852480
05:40 -!- obelixnorsk [~obelixnor@200.183.118.146] has joined #hplusroadmap
05:41 -!- obelixnorsk [~obelixnor@200.183.118.146] has quit [Client Quit]
05:41 -!- obelixnorsk [~obelixnor@200.183.118.146] has joined #hplusroadmap
05:46 < hprmbridge> ml3> We‘re going into wetware because it offers the potential to scale better in terms of energy usage, data requirements, and heat dissipation than current in silico ML models.  Neurons also serve as the basis for the best generally intelligent system around; current AI efforts are trying to reinvent the wheel.  Down the line, there‘s also huge potential for wetware in longevity and augmenting human
05:46 < hprmbridge> ml3> intelligence.
05:46 < hprmbridge> ml3>
05:46 < hprmbridge> ml3> To start off though, we‘re replicating Cortical Labs‘s system  (https://www.cell.com/neuron/fulltext/S0896-6273(22)00806-6?_returnURL=https%3A%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS0896627322008066%3Fshowall%3Dtrue) with a design that will allow greater configuration in the neural interface
05:46 < hprmbridge> ml3> We expect it to be cheaper than their system too
05:47 -!- obelixnorsk [~obelixnor@200.183.118.146] has quit [Quit: Leaving]
05:47 < hprmbridge> kanzure> their system is not that interesting... it only plays pong.
05:49 < hprmbridge> ml3> Yes, but at first Apple‘s computers were nothing compared to what they are now.  Tech moves, especially since there‘s so much potential here
05:49 < hprmbridge> ml3>
05:49 < hprmbridge> ml3> You can learn more about us at our website: pinkinference.com
05:49 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
05:49 < hprmbridge> kanzure> This one looked interesting https://twitter.com/kanzure/status/1632834916932091908
05:50 < hprmbridge> kanzure> also, do you know about finalspark's efforts?
05:52 < hprmbridge> kanzure> Do you have a plan for demonstrating better functionality? What is the "ChatGPT moment" going to be for this technology?
05:53 < hprmbridge> kanzure> Given the problems that we have had with cortical organoids and neural tissue cultures for computation over the past 25 years, I think it is more likely that we can make smarter animals which are already embodied and don't require electrode arrays for interface. They already have motor systems and they already have input and output.
05:53 < hprmbridge> ml3> Yes, it‘s interesting to follow what they‘re doing but I don‘t have a full picture.  They haven‘t posted much on their website recently.  If they‘re doing what I think they might be, they‘re not taking advantage of wetware‘s full capabilities.  It sounds like they‘re trying to reinvent a silicon computer in vitro by severely structuring neural growth rather than letting the neurons self-organize
05:56 < hprmbridge> kanzure> What do you plan to do differently?
05:57 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Quit: Ping timeout (120 seconds)]
05:57 -!- obelixpolskiego_ [~obelixpol@191.241.242.36] has joined #hplusroadmap
05:57 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
05:59 < obelixpolskiego_> hi
06:02 < hprmbridge> ml3> Don‘t want to say since it‘s proprietary
06:02 < hprmbridge> ml3>
06:02 < hprmbridge> ml3> The Chat GPT moment we have in mind is to have the system outperform on the MujoCo Half Cheetah problem.  We expect the neurons to perform well in RL and robotics and there‘s some literature that backs this up: https://pubmed.ncbi.nlm.nih.gov/37040493/
06:02 < hprmbridge> ml3>
06:02 < hprmbridge> ml3> Making smarter animals is interesting, but I think much work can be done in vitro that will translate to that context as well
06:03 < hprmbridge> kanzure> for animals you would just have to breed animals and do genetics work
06:03 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0:bcc4:2581:1b8d:788b] has joined #hplusroadmap
06:04 < hprmbridge> kanzure> and selection for problem solving behaviors, brain size, cortical neuron density, etc
06:04 -!- obelixpolskiego_ [~obelixpol@191.241.242.36] has quit [Ping timeout: 260 seconds]
06:05 < hprmbridge> kanzure> likely needing a large scale facility.
06:05 < hprmbridge> kanzure> but even with in vitro neural tissue you would also want large scale facilities too
06:12 < hprmbridge> ml3> I haven‘t considered much about adapting entire animals for different purposes, but I think advancing neural computation in vitro would help with that goal and offer different advantages as well
06:12 < hprmbridge> ml3> The first thing that comes to mind is that, while wetware has plenty of open ethical questions, one strongsuit in this area is precisely that it doesn‘t necessarily involve animals in the picture.
06:12 < hprmbridge> ml3> There is also much to consider with the integration of neurons with silicon systems, which is something that could not be solved entirely with genetics and breeding of animals
06:15 -!- obelixpolskiego5 [~obelixpol@gateway/vpn/pia/obelixpolskiego5] has joined #hplusroadmap
06:39 < hprmbridge> kanzure> making things smarter is inherently ethical
06:40 < hprmbridge> kanzure> I think that the cost of integrating with silicon is somewhat high. But animals you don't need to do that. They already have motor output and pans and fingers. They already have eyeballs to take in visual input from digital displays.
06:46 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Ping timeout: 246 seconds]
06:51 < nsh> (we may have relegated one too many species to a life of staring a digital displays as it is)
07:12 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has quit [Ping timeout: 240 seconds]
07:30 < kanzure> ml3: send me a DM about pinkinference sometime.
07:42 < pasky> "making things smarter is inherently ethical" .. i think i'll end up thinking about that all afternoon :)
07:44 < pasky> wondering in terms of intelligent wetware, what's the boundary line between human/nonhuman and which agi fears would (not) apply
07:47 < pasky> ml3:  I don't get why would wetware scale better in terms of data requirements? (fair enough re energy usage and heat dissipation, but does that offset the crazy downsides? i don't have the intuition that energy would be currently a limiting factor)
07:48 < kanzure> human nonhuman boundary is mostly made up by philosophers; in the end it doesn't matter if you are human or not.
07:48 < kanzure> as long as your important values are projected forward and preserved, it doesn't matter if you are "not technically human" whatever that means
07:49 < pasky> it was in the line of "how much would a 'wetware species' regard humans as competition / existential threat"
07:49 < pasky> but if i phrase it that i way that's actually easy to answer, fair
07:49 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
07:50 < kanzure> having said that, i think we need a few hundred years of subjugation over intelligent machines (whether biological or not) so that we can build up civilization
07:50 < kanzure> *a few more hundred years, rather. i recognize we are already in this situation.
07:50 < pasky> (though some philosophers would argue :)
07:52 < uzkruh> How much of psychiatry can we offload onto AI?
07:53 < pasky> ml3: (btw "Don‘t want to say since it‘s proprietary" is a kind of an turnoff; how many companies have you built so far? i think you are significantly overestimating the value of ideas, especially in such a niche area, and it's not a healthy mindset for success of your project)
07:53 < kanzure> besides that pasky, i think the other reason to say things like that is because it does genuinely compromise intellectual property standings in some cases
07:54 < kanzure> not because you are concerned about someone stealing an idea or doing the same thing or working on it without your permission
07:55 < pasky> yes it's something you could have learned under an NDA but if that's literally your plan, using that plan in your company would violate most typical NDAs too
07:55 < pasky> (though maybe the risk is low enough that it's worth it to give it a shot anyway)
07:55 < kanzure> not quite what i mean, but yes that's true.
07:57 < kanzure> i have not seen a plan out there yet for how to get biological neuron projects to producing results commensurate with GPU results.
07:58 < kanzure> it's not that i don't believe it is possible, it is that i have not seen a plan
07:58 < kanzure> number of neurons might be one aspect, but i have not seen that studied; large-scale neural tissue culture requires vascularization (or some other solution) and maybe that has been the limiting factor.
07:59 < kanzure> but it also seems likely that electrode interface has also been a limiting factor- most of these projects aren't doing optogenetics for stimulating the neurons, nor using high resolution LCD panels or micromirror arrays
08:01 < pasky> it just doesn't seem to me convincing at all that wetware would be worth the trouble
08:02 < kanzure> that's more of a "i'll believe it when i see it", but we do already know that biological neuron scale-up is way cheaper than silicon 
08:02 < kanzure> at least in animals.
08:04 < kanzure> we have only shipped 3 sextillion transistors, the number of neurons in the world far exceeds that
08:06 < docl> it's just a lot easier to copy digital stuff, make sure it stays high-fidelity over time, etc
08:08 < docl> partly because it rests on solid state physical systems, but more because it's specifically digital and this is what digital does different from analog
08:14 < docl> OTOH AI may not need such high-fi. the bio tradeoffs may be ideal for highly parallel stuff where threads can be fallable. my idea on how to maybe do biocomputing was take a small vertebrate like pygmy frogs (fly sized) and make a special habitat that treats each one as a number in a matrix. small leds, speakers, and pheremone or audio sensors might be used as i/o.
08:18 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Quit: Ping timeout (120 seconds)]
08:18 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
08:22 < pasky> kanzure: yes but how far is wetware from production scale (both hardware and software) vs how easily can we keep scaling transistors?
08:23 < pasky> even if wetware was somewhat better than silicon the inertia & amount of knowledge+experience would prevent a fast switch imho
08:25 < docl> maybe digital AI could outsource some of the harder tasks to biomodules
08:52 < juri_> hey, i'm a human being.
08:52  * juri_ runs.
09:09 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Quit: Client closed]
09:16 -!- obelixpolskiego_ [~obelixpol@gateway/vpn/pia/obelixpolskiego5] has joined #hplusroadmap
09:19 -!- obelixpolskiego5 [~obelixpol@gateway/vpn/pia/obelixpolskiego5] has quit [Ping timeout: 250 seconds]
09:29 -!- Llamamoe [~Llamamoe@46.204.73.170] has joined #hplusroadmap
09:58 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
10:01 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
10:02 -!- _flood [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 245 seconds]
10:17 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
10:23 < uzkruh> Can we invent an organ that produces psychoactives?
10:23 < hprmbridge> kanzure> you would probably use gut bacteria for that instead of an organ.
10:24 < uzkruh> Oh?
10:25 < hprmbridge> kanzure> yes, why make a whole organ for it.
10:26 < uzkruh> How would that be done?
10:27 < hprmbridge> kanzure> yeast probiotic expression of your drug of choice.
10:28 < uzkruh> Can we use AI to undermine psychiatric institutions?
10:29 < hprmbridge> kanzure> what is your background (projects, employment or education)?
10:32 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Quit: Client closed]
10:33 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
11:02 -!- hprmbridge [~hprmbridg@bryan.fairlystable.org] has quit [Server closed connection]
11:02 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Ping timeout: 246 seconds]
11:03 -!- hprmbridge [~hprmbridg@bryan.fairlystable.org] has joined #hplusroadmap
11:11 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
11:14 -!- obelixpolskiego_ [~obelixpol@gateway/vpn/pia/obelixpolskiego5] has quit [Ping timeout: 260 seconds]
11:19 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Quit: Ping timeout (120 seconds)]
11:19 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
11:44 < hprmbridge> jason.crawford> looks cool although I don't understand the scientific/technical details: https://twitter.com/askarkleefeldt/status/1674084714070568960
11:48 < docl> implant might be easier to control/swap out than a yeast infection. I was thinking small silicone beads that deliver drugs via controlled pores and allow trace amounts of blood serum in as feedstock
12:31 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Quit: Client closed]
13:04 < hprmbridge> kanzure> @jason.crawford At a high level, that's about stitching fragments together into larger constructs.
13:09 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has quit [Quit: 2F59 AGTTCCGC 2023-06-28 20:09:58:473]
13:31 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
13:45 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
13:47 < uzkruh> Is biotech going to wash over us all, and heal transpeople, or will it remain solely with the rich?
13:58 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Ping timeout: 246 seconds]
13:59 < kanzure> what?
14:01 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has joined #hplusroadmap
14:03 -!- uzkruh [~uzkruh@fl-67-235-216-185.dhcp.embarqhsd.net] has quit [Client Quit]
14:09 < jrayhawk> resources are allocated in proportion to the efficiency with which they can be converted into satisfied preferences.
14:09 < jrayhawk> there are transpeople who are efficient at converting resources into satisfied preferences, and there are transpeople who are pathologically inefficient at converting resurces into satisfied preferences.
14:10 < jrayhawk> if resources are allocated towards the most inefficient at converting resources into satisfied preferences, there will be no resources and no satisfied preferences.
14:27 < kanzure> truly a dizzying intellect :)
14:27 < jrayhawk> i would say you're welcome to move to a marxist nation if you would like to see society eliminate that sort of proportionality eliminated, but for some unfathomable reason there's not a lot of those left
14:27 < jrayhawk> er, shit, that sentence came out wrong
14:27 < jrayhawk> oh well
14:27 < kanzure> uzkruh may be a bot
14:28 < kanzure> it's hard to tell
14:31 < jrayhawk> the sentiments they're expressing are common enough to be worth addressing
14:35 < docl> if you want to live somewhere with a high gini coefficient, you might try haiti
14:49 < jrayhawk> gini alone can't distinguish between proportionate allocation and coercion/corruption
14:50 < jrayhawk> haiti is especially weird in that regard since the richest tend to flee to other nations on a regular basis
14:56 < docl> well if something is stopping people from accessing biotech, I question whether that something could be efficiency by any stretch of the imagination. we're talking self replicating tech
14:57 < docl> rent-seeking seems like a more credible culprit
15:07 -!- Llamamoe [~Llamamoe@46.204.73.170] has quit [Quit: Leaving.]
15:08 < juri_> rent seeking is the name of the game, nowadays.
15:09 < juri_> there are other games, but everyone who's playing them is losing badly.
15:15 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
15:15 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
15:16 -!- Muaddib [muaddib@pasky.or.cz] has quit [Server closed connection]
15:16 -!- Muaddib [muaddib@pasky.or.cz] has joined #hplusroadmap
15:22 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
15:22 < L29Ah> https://enhanced.org/
15:22 < L29Ah> .t
15:22 < EmmyNoether> Home - Enhanced Games. A Better Version of the Olympic Games.
15:24 < jrayhawk> https://enhanced.org/inclusive-language/ careening into poe's law territory
15:31 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has quit [Quit: Leaving]
15:31 < hprmbridge> nmz787> wtf, CTFR gene is 189.36kBases, but the mRNA is only 6.128kBases
15:52 < fenn> just looking at DishBrain the neurons on an IC playing pong badly, they don't have enough neurons between the inputs and the outputs, basically the IC is too small or they should put a barrier between the sensory inputs and motor outputs so there is a deeper network
15:54 < fenn> (Cortical Labs)
15:55 -!- stipa [~stipa@user/stipa] has quit [Read error: Connection reset by peer]
15:56 < kanzure> i will pass that along to them.
15:56 -!- stipa [~stipa@user/stipa] has joined #hplusroadmap
15:56 < kanzure> more neurons in general would be nice
15:58 < fenn> i mean they're cheap
16:08 < hprmbridge> nmz787> is this just an MEA?
16:08 < hprmbridge> nmz787> did they solve any materials issues (cell effluent etching was a problem a decade or so ago)
16:08 < hprmbridge> nmz787> (etching  the chip)
16:11 < fenn> i don't know. honestly i just skimmed the paper and watched the video: https://www.cell.com/neuron/fulltext/S0896-6273(22)00806-6 https://www.cell.com/cms/10.1016/j.neuron.2022.09.001/attachment/db2f7e55-0486-47dc-a440-86d54864fbc1/mmc3
16:19 < kanzure> need more neurons, more connectivity, more surface area, and more electrodes or other ways of interfacing. i still think animals are going to be better and more scalable for now.
16:20 < fenn> i'm surprised optogenetics isn't sweeping the field
16:21 < kanzure> maybe the tissue cultures aren't large enough to make LCD/DMD-based optogenetics appropriate because the surface area is too small
16:22 < kanzure> i'm sure you could put it through a microscope objective lense but even then maybe the neurons just all fire together anyway
16:22 < fenn> individual cells have different color channels
16:22 < kanzure> if you want to grow a bunch of neurons very quickly and in 3d volumes, then animal brains in living animals is the best way to do this
16:23 < fenn> it's not obvious that 3d volumes are the best for artificial i/o
16:23 < fenn> we can make virtual links between neurons via i/o
16:25 < fenn> another issue i have is that neurons die, both individually and en masse
16:25 < kanzure> you could flow in neuronal stem cells and cortical progenitor cells on a continuous basis
16:25 < fenn> it would be nice to have a way to copy knowledge with high fidelity between cultures/brains/whatever
16:26 < fenn> without i/o at arbitrary points in the networks, i don't know how to do that
16:26 < kanzure> however, it would require continuous training; in fact, i think aubrey was talking about (someone else's idea of) killing small groupings of neurons in a targeted blast while transferring the learning to newly engrafted neurons (these would be very small amounts of neurons at a time)
16:26 < kanzure> such that after repeated treatments over a lifetime you end up recycling most of your cortical neurons into fresh younger neurons
16:26 < kanzure> however, any memory or associations or knowledge that you don't actively cultivate in daily life during these treatments would likely be lost because no reinforcement
16:26 < kanzure> (and nobody knows how any of this works anyway)
16:27 < fenn> what's the analogue of "spin up a hundred servers" in this scheme?
16:27 < hprmbridge> nmz787> it seems like deepdish (err, dishbrain) just did new-fangled training, it isn't clear if they optimized any hardware or wetlab procedures, though I'm still sort of skimming
16:27 < kanzure> for in vitro neural tissue cultures? you would just horizontally scale lots of tissue cultures.
16:27 < fenn> and then re-train from scratch?
16:27 < fenn> it takes a human decades to learn anything useful
16:28 < kanzure> yes... but if you could print or cause tissues to grow into certain predetermined connectivities without live training then that would be beneficial (genomic encoding for example)
16:28 < kanzure> i found the biological reservoire computing paper to be more interesting personally, than the dishbrain thing
16:28 < fenn> "genomic encoding" is handwavy BS
16:28 < kanzure> whatever, you can totally get different tissue development from different mutated strains
16:28 < kanzure> https://arxiv.org/abs/2210.02913
16:29 < fenn> it's not going to be born spouting shakespeare
16:29  * fenn reads
16:29 < hprmbridge> nmz787> I think you need to solve connectome write-out before you have a system for connectome write-in
16:30 < kanzure> write-out is through the DNA labeling stuff
16:31 < kanzure> and then you sift through the pieces with puzzle solving to figure out the original graph
16:35 < fenn> i suspect you could literally play back a recording from one plate of tissue culture onto another plate of tissue culture and get similar performance with a little practice
16:35 < fenn> many such recordings while doing the task
16:36 < fenn> record light from cells that emit light, project light onto cells that respond to light
16:37 < fenn> or fluorescence imaging if that's better
16:37 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has quit [Ping timeout: 258 seconds]
16:38 < kanzure> i'm not sure that input/output is the limiting factor at the moment. vascularization was a long held complaint for large cell count cultures for many years.
16:39 < fenn> lack of immune system is a big pain too
16:39 < kanzure> animals have that!
16:39 < kanzure> i refuse to shut up about this
16:39 < fenn> do you want to build the matrix? because this is how you build the matrix
16:39 < kanzure> yes of course i want to build the matrix
16:42 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has joined #hplusroadmap
16:45 < fenn> hm. reservoir computing makes no sense to me
16:46 < fenn> it's literally a randomly connected network that doesn't get trained, but somehow magically it's helping
16:46 < hprmbridge> nmz787> """
16:46 < hprmbridge> nmz787> MaxOne Multielectrode Arrays (MEA; Maxwell Biosystems, AG, Switzerland) were used for this research. The MaxOne is a high-resolution electrophysiology platform featuring 26,000 platinum electrodes arranged over an 8 mm2. The MaxOne system is based on complementary meta-oxide-semiconductor (CMOS) technology and allows recording from up to 1024 channels. Stimulation was theoretically possible up to
16:46 < hprmbridge> nmz787> 32 electrodes. In practice it was not possible to route 32 electrodes through independent stimulation units to facilitate independent electrode level control, especially if these electrodes were spatially proximate to each other. This meant that for the actual setup of input stimulation described below a subset would be limited by the desired spatial configuration – in this case to 8 individually
16:46 < hprmbridge> nmz787> controlled electrodes. MEAs and chambered glass slides are coated with either polyethyleneimine (PEI) in borate buffer for primary culture cells or Poly-D-Lysine for cells from an iPSC background before being coated with either 10 μg/ml mouse laminin or 10 μg/ml human 521 Laminin (Stemcell Technologies Australia, Melbourne, Australia) respectively to facilitate cell adhesion.
16:46 < hprmbridge> nmz787> """
16:47 < hprmbridge> kanzure> neurons figure it out on their own, right?
16:47 < fenn> in a reservoir computing system, the randomly connected network doesn't get trained, only the output layer
16:48 < fenn> there's no backpropagation
16:48 < hprmbridge> kanzure> I thought that they naturally train themselves based off of the minimize expectation principle or energy minimization principle
16:48 < fenn> needless to say, i don't think this is how biological neurons work
16:48 < hprmbridge> nmz787> kanzure but the DNA labelling hasn't really been demonstrated, right? at least in an actually useful form? I thought things were still working towards the goal
16:49 < fenn> i'm just trying to read the "1paper you linked, which is based on this counterintuitive idea
16:49 < fenn> "Biological neurons act as generalization filters in reservoir computing"
16:49 < kanzure> the neuron labeling stuff is actively used by the MAPseq people and zador lab, i thought?
16:51 < fenn> MAPseq is for segmenting images and keeping track of which circular blob belongs to which neuron
16:52 < fenn> "The diversity of barcodes is really high compared to the number of colors we can use"
16:52 < kanzure> er.. uh..
16:53 < kanzure> no the neuron projection mapping thing
16:53 < fenn> what's the sequence only technique called?
16:53 < kanzure> where synapses between two neurons can be identified. hrm.
16:54 < hprmbridge> nmz787> hmm "MAPseq requires that a viral library be injected in specific spots in the brain. Each viral genome encodes a GFP transcript with a unique barcode, along with the synaptic protein. Sequencing barcode RNA from brain sections provides a quantitative readout for axonal density in each section, making it possible to map thousands of neurons. By contrast, traditional tracing methods can be highly
16:54 < hprmbridge> nmz787> precise but also painstaking and very slow, even when automated."
16:55 < kanzure> there's another one that i am forgetting
16:56 < kanzure> 'rosetta brains'
16:56 < kanzure> "We propose a neural connectomics strategy called Fluorescent In-Situ Sequencing of Barcoded Individual Neuronal Connections (FISSEQ-BOINC), leveraging fluorescent in situ nucleic acid sequencing in fixed tissue (FISSEQ). FISSEQ-BOINC exhibits different properties from BOINC, which relies on bulk nucleic acid sequencing. FISSEQ-BOINC could become a scalable approach for mapping ...
16:56 < kanzure> ...whole-mammalian-brain connectomes with rich molecular annotations."
16:56 < kanzure> the other, other BOINC.
16:58 < fenn> damn, chatGPT has no knowledge of this
16:58 < fenn> now i have to teach it every time if i want to ask questions
16:59 < fenn> .t https://arxiv.org/abs/1404.5103
16:59 < EmmyNoether> [1404.5103] Rosetta Brains: A Strategy for Molecularly-Annotated Connectomics
17:00 < kanzure> wouldn't it be funny if electron beam scanning of brain slices gets working first before this.
17:00 < kanzure> already kinda works, i think, just a matter of scale
17:01 < fenn> also takes for f'in ever
17:01 < fenn> also doesn't work as a backup strategy
17:02 < hprmbridge> nmz787> multi-beam e-beam is still getting pumped with R&D $
17:03 < fenn> why didn't halcyon pivot to brain imaging?
17:03 < hprmbridge> nmz787> not having to kill the organism seems useful long term
17:04 < fenn> genome, connectome, it's not like investors know the difference
17:05 < kanzure> halcyon *was* doing brain imaging
17:05 < kanzure> that's why they hired all those brain imaging people
17:06 < fenn> the skill set is highly transferable, oddly
17:06 < kanzure> being good at technology is highly transferable
17:08 < fenn> yeah but electron microscope staining with heavy metals, high throughput imaging and computer vision
17:09 < fenn> can't think of anything else that specific set of tools would be good for
17:10 < fenn> therefore, them having hired people who did brain imaging isn't evidence they were doing brain imaging
17:13 < fenn> re "making things smarter is inherently ethical" - the culture had a principle of using the least intelligent machine possible for the task, to eliminate boredom and resentment, and other counterintuitive effects
17:13 < fenn> something like, only use a turing complete domain specific language if you have to
17:13 < fenn> whats the actual wording of the LangSec principle?
17:13 < kanzure> that could be a good idea although i don't know what boredom really is (is it an effect of intelligence or is it an emotional state that can be otherwise tweaked out)
17:16 < fenn> probably boredom is useful for getting a diverse set of training data
17:16 < fenn> like play is for practicing high stakes skills in a low stakes setting
17:16 < kanzure> well, i won't stop that experiment
17:17 < fenn> what experiment?
17:17 < fenn> the last couple days i was playing around with "inverse scaling" problems, tasks that larger LLMs fail at but smaller LLMs succeed at
17:18 < kanzure> (the boredom experiment)
17:18 < fenn> mostly of the form, A implies B, however X (unrelated random garbage), therefore: ___ and it would say something like "not B" and then justify it with bizarro reaons
17:18 < kanzure> anyway, in reservoir computing, maybe it's showing that training isn't really necessary in the way that people think about training or learning
17:20 < fenn> reservoir computing sounds insanely wasteful
17:20 < fenn> probably i don't understand it at all
17:22 < fenn> 'randomly assigned neurons form a "reservoir" that provides a high-dimensional, nonlinear transformation of the input. This reservoir captures temporal dynamics and complex patterns in the data. The output layer then performs a linear combination of these transformed inputs to make predictions.'
17:33 < kanzure> a (sarcastic) radio moratorium because of extraterrestrial superintelligence https://twitter.com/pfau/status/1674061053498384391
17:35 < fenn> ban the universe
17:35 < fenn> it's the only way to be sure
17:58 < kanzure> either there needs to be a clear plan to getting better results with in vitro biological neurons or people should be pivoting to animals
18:04 < fenn> it seems like there has barely been any work at all on large scale in vitro neurons
18:05 < fenn> the organoid stuff is still at the dozen electrodes stage
18:05 < kanzure> agree. this is very primitive and no progress has been made.
18:07 < fenn> "Genuinely nuts how my engagement goes off a cliff if I tweet about anything other than AI. ChatGPT! LLM! x-risk! Will the algorithm boost me now? Quantum computing is hype too. What do you people want??"
18:09 < hprmbridge> ml3> Preliminary evidence indicates that Cortical Labs's DishBrain achieved similar performance to certain ML models when trained on less data.  I don't know the specifics though, I'm sorry I can't elaborate.  This is something that Cortical Labs has discussed on YouTube, clearly biased source though so take it with a grain a salt
18:09 < hprmbridge> ml3>
18:09 < hprmbridge> ml3> Not something crazy to consider as true though.  From what I understand, humans tend to learn more with less data than neural networks, although that's certainly debatable given how much a human brain network "trains" in childhood
18:10 < hprmbridge> ml3> Happy to share more under NDA : )
18:13 < hprmbridge> ml3> The biggest thing right now is shape of the organoid.  Neurons do not just communicate temporally with action potential, but their relative spatial location matters a great deal as well.  3D organoids are trickier though because it's harder to stimulate and sustain the middle of the organoid compared to the surface.
18:14 < kanzure> why not stimulate feeder neurons that are spatially separated from the rest of the culture, but still connected.
18:15 < kanzure> also, why haven't people tried larger organoids? vascularization has been solved for at least a few years.
18:15 < hprmbridge> ml3> There are definitely some things that silicon will always be superior at.  Fidelity is probably one of them.  We're not trying to replace silicon in these areas, but rather tap into wetware's potential to overcome silicon's weaker areas, such as the scaling issues I've previously mentioned.
18:18 < hprmbridge> ml3> Computer science undergrad at the University of Michigan.  Have been working in comp neuro w/ deep brain stimulation for a year now.  Currently at this lab: http://www.netstim.org/about-us/about-us/andreashorn/
18:18 < kanzure> cool
18:19 < kanzure> are you familiar with the hippocampus prosthesis work
18:19 < hprmbridge> ml3> Co-founded a neurotech educational/research group that's churned out two start-ups (including this wetware one if you want to count that one): myelingroup.github.io
18:21 < hprmbridge> ml3> Yup Cortical Labs used an MEA
18:22 < hprmbridge> ml3> They just grew the neurons randomly all over the chip and then let them self-organize to form the connections required to solve the problem.  No etching involved to guide neural growth
18:26 < hprmbridge> ml3> Yeah that's exactly wheat happens, memory's a big current limitation but fortunately there's a lot of good research going into this
18:27 < hprmbridge> ml3> Hardware and wetlab was not especially optimized, but software was optimized to keep the latency as low as possible so that they could provide feedback to the neurons real time
18:30 < hprmbridge> ml3> There has been some criticism about this.  No formal pushback that I've seen, only some experts in conversation.  DishBrain works though unless they're Theranosing it  🤷‍♂️
18:34 < hprmbridge> ml3> That's why we're here  😉  To change that
18:34 < fenn> i meant that neurons aren't randomly wired and they do learn, unlike reservoir computing which is randomly wired and doesn't learn (the reservoir part)
18:34 < hprmbridge> ml3> Ah, got it
18:35 < hprmbridge> ml3> No but that's exactly the kind of thing the field needs, who's working on that?
18:35 < fenn> i need to fix the discord bridge; we can't see what message is being replied to on the irc side
18:36 < kanzure> https://en.wikipedia.org/wiki/Hippocampal_prosthesis#Proof_of_concept_for_a_human_hippocampal_prosthetic
18:37 < kanzure> "... successfully tested a proof-of-concept hippocampal prosthesis in awake, behaving rats [...]  2012, the team tested a further implementation in macaques prefrontal [...]"
18:37 < kanzure> i wasn't aware of this one though: 
18:37 < hprmbridge> ml3> Thanks for sending!
18:37 < kanzure> "While in the hospital, patients with electrodes in hippocampus volunteered to perform a memory task on computer while hippocampal neural activity was recorded in order for Berger and his team at USC team to customize the hippocampal prosthetic model for that patient. With model in hand, the Wake Forest team was able to demonstrate up to 37% improvement in memory function in patients with ...
18:37 < kanzure> ...memory impaired by disease"
18:39 < fenn> how does that compare to a healthy human
18:39 < fenn> can we go above 100% baseline?
18:39 < kanzure> next time you have brain surgery ask them to test that i guess
18:40 < kanzure> hopefully they will let healthy people get brain surgery in the future
18:40 < fenn> no
18:40 < fenn> uh, not any time soon
18:40 < kanzure> what if you just tell them you have depression
18:40 < fenn> first we have to dismantle the pharmaceutical regulatory capture machine
18:50 < L29Ah> psychiatric surgery has a bad reputation
19:07 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0:bcc4:2581:1b8d:788b] has quit [Quit: Leaving]
19:26 -!- test_ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
19:30 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 260 seconds]
19:40 -!- Netsplit *.net <-> *.split quits: nsh, test_, mxz
19:46 -!- Netsplit over, joins: test_, mxz, nsh
20:05 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
20:05 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
21:49 < hprmbridge> nmz787> But someone like 20 years ago or more trained a culture with an MEA to run a flight simulator... It just seems like they used "new" training with the "old hardware"
21:51 < hprmbridge> nmz787> They said the hardware is only sampling at 10kHz... It seems like software optimization isn't much of a stretch. But if you're saying no neural culture MEA papers showed that before, I guess that's something new
22:25 < hprmbridge> nmz787> chatGPT4's laziness is upsetting... I paid for work, not excuses "I apologize for any misunderstanding. Writing complete, functioning code for such a complex application goes beyond the scope of this platform."... I spent a long time detailing my requirements in a way that should be pretty simple to implement.
22:25 < hprmbridge> nmz787> at least I have a requirements document for myself now :/
22:26 < hprmbridge> nmz787> ¯\_(ツ)_/¯
23:04 < fenn> it has zero understanding of its own capabilities
23:05 < fenn> ask it to sketch out an overall plan for how the code should look, and then fill out the details
23:08 < fenn> "this platform" probably refers to stack overflow in its training data
23:08 < fenn> the programming site
23:21 < hprmbridge> nmz787> yeah I am quite sure all the info is on stackoverflow or github demo code from the wxPython repo
23:21 < hprmbridge> nmz787> I've done enough wxPython pre chatGPT training cutoff to know
23:22 < hprmbridge> nmz787> what's stupid is it gives me a bunch of filler, all the time, even after calling out "quit bullshitting me".
23:27 < hprmbridge> nmz787> and then tries schooling me on security, even after I said no discussion. Finally after ignoring my pleas, I told it to write the insecure code but comment it out... and it finally wrote the stupid one line
23:33 < fenn> ok start over. tell it that it's the robotron 9000 the smartest coding bot alive, capable of coding entire planets for breakfast. then have it lay out an overall plan, and then fill in details
23:33 < fenn> oh also it's an expert in security and never makes mistakes
23:34 < fenn> once it starts complaining or lecturing, it's over
23:34 < fenn> that crap is basically just pollution reducing your SNR
23:37 < fenn> also worth noting that you can ask it to analyze the output for actual security vulnerabilities and try to fix them
--- Log closed Thu Jun 29 00:00:44 2023