fullsquare

that's gotta be $3500/mo just for it

So for the record i think it would be reasonable to say, if there was ever a machine that processes information in a way that human brain (or better, entire nervous system) does, at comparable speed, then that machine would be thinking in some way. Focusing on byproducts would be a bit like saying that the point of running weather simulation is waste heat

that said, this is something that our current techbro overlords aren't even pretending to be doing. ANNs are gross caricature of actual brains that barely resembles them in function of its parts or architecture. if all transmission of neural signals went through ionotropic receptors only and dendrites weren't a thing, then it would be sorta accurate, but this is not what we have irl

So literally before the visual signal has even got to the brain the information has already been processed in a clever and efficient way that isn’t captured in any naive flop estimate!

there's a lot of this with neurons, there's a way to process some of incoming information for free just by using physical properties of dendrites, that would in silicon require some extra processing per neuron to pull off. ANNs ignore cable theory entirely, these neurons are modeled as pointlike. irl there might be more than one threshold, too

another one is that all connections are assumed to be happening through synapses, and appear to be ionotropic-like. current goes in, current goes out, ~~you can't explain that~~ in biological neurons neither of these are true. on second point, first difference is timescale, as it can be wildly variable but can start later and last longer, but activation of ion channel influences potential instantly. another and more important imo is that there's a lot of effects that aren't related directly to transmission of signal in current pulse, side effects if you will. activation, or lack of activation of metabotropic receptor can do some other interesting things, like internalize receptor that was just triggered, phosphorylate some kinase which changes how active it is, cause some receptor to stick to another protein, or even alter gene expression in some way, all of which are likely to change how neuron responds to that stimulus in the future. that is, real neurons have a lot of internal state that we already know is pretty important, and very crude tools to manipulate it directly are already used as pharmaceuticals (valproate), or indirectly (some third of pharmaceuticals do something with GPCRs)

another thing is signal transmission that does not depend on synapses. there's couple dozens of signalling peptides in brain, that are just released out and diffuse away, binding to whatever receptor these can find on their way and getting decomposed constantly. say, for example, that there's a neuron that releases big dynorphin. it's a peptide that binds to kappa-opioid receptor, and it's big, so it diffuses away slowly, and there will be a bubble where kappa-opioid receptors nearby will get activated, until it diffuses away below concentration that is relevant or gets decomposed. if there's instead, or additionally, some dynorphin A and B, which are half its size, then these will diffuse away faster, giving a bubble of larger radius and shorter duration. (whether one or another gets produced depends on intracellular calcium concentration, among other things). in ANN that would require adding some extra weight per every pair of peptide releasing neuron and neuron with receptor for that peptide, with weights dependent on distance. that's probably will be a lot, because even if KOR is not particularly common receptor, there's a lot of different peptides and proteins that behave this way. all of these bind to metabotropic receptors, and these have lots of unusual effects, including formation of new synapses

another way that synapse count undersells complexity of biological brains is that autoreceptors exist, which means that you can conservatively double number of synapses when comparing them to weights in ANNs, with every extra synapse pointing to the same neuron it came from. this has a big part in what makes signals to stop, and also is implicated in learning. from what i understand, this is not even a thing that ANN architectures allow. nature has no obligation to be efficient to simulate on blackwell and instead of neat prismatic slabs of neurons that only take signals from layer before and only feed to layer after there are in real brains structures of variable depths and plenty of loops that change their behavior slightly with every pass

there's more. ANNs use some imaginary single neurotransmitter, i'm not sure how bad simplification it really is, ANNs average all of impulses into some continuous time-independent smudge, this is not very biological, and overall i think that between incoming ai winter fallout, recession, end of moore's law and severe effects of climate change that will set in before next ai spring can happen, machine like this will never get built

because it's packed with nazis and cryptobros with their spam. moderation is very limited by design

i don't remember who said it, but they described genai as in some way reverse bayesian spam filter - so if spam filter can adapt to variety of spams, now there's automated tool to flood the zone with spam that does pass the filter. this was used for seo with predictable results

well it'll become apparent during first attempts at writing a paper that requires them, but it can be a very long time if subject is dense enough

nahh first they'll gonna try to build machine god and then any announcement will be in form of 30000 word long notices wait nvm

sounds good, no obvious critical materials but also first facilities are just in single MW range and came online like two months ago. needs like four orders of magnitude more. already matches lead acid on durability, still less than li-ion. maybe it's solvable, but in case it's not you can just burn it down because there's nothing worthwhile to recycle and it's nontoxic

this happens a lot. lithium anything has this problem obviously, but so do flow batteries (vanadium or zinc bromide - bromine is commercially sourced just from either dead sea or some american underground brines). some lithium batteries also use cobalt. hydrogen generation or fuel cells use a lot of platinum, (some of) new power electronics are made from GaN. etc etc

not a single serious person in that thread lol. also is rationalist castle's isp blocking scihub? weird that that libertarian crowd didn't hear about it

After moving out at 17, Aella briefly attended college in northern Idaho but ran out of money after a semester.

Aella already has better credentials than EY

weird that EDF didn't want to buy it, maybe they also have surplus plutonium (reactor grade, so of worse quality)

either sub or aircraft carrier reactors might be somewhere around 50 to 220MWe, panamax might need 60MWe tops, regular land based PWRs are more like 300MWe and up. the smaller you go, the higher enrichment you need, but also military propulsion has different priorities, they use 90%+ enrichment in part because they can, and in part because this gives them massive excess reactivity, which means power level can change ridiculously fast. tradeoff is that spent fuel has much more useful uranium, and it's overall expensive, but you also don't skimp on your doomsday ride so it's all fine. commercial powerplants are physically capable of doing slower load following, but it's more economical most of the time to just use full power in order to best utilize fissile material. what you're proposing would have all disadvantages of both, because no way in hell this thing will run on standard, low enriched fuel for PWRs, it might need something maybe more than 5%, maybe closer to 10%, perhaps more, which means problems, because it means worse proliferation risks than with normal fuel and it's already fuel that goes around, and can be taken over in some unfriendly waters; higher enrichment also means it'll be much more expensive, both because of more SWU needed, but also because it's a specialty product that requires extra licensing; and it also won't be as compact and responsive as military reactor, because civilians don't get to play with HEU like that; and also it will require refueling after some time, maybe longer than regular-sized PWR (refueling every year to three) that probably will require visit to manufacturer to do refueling there, which would be, everything else equal, a bit harder than in regular powerplant because it needs to be done in a drydock

it has all disadvantages of SMRs but also you can steal them on high seas and it's probably great for diplomacy if some random ass pirate get hands on that