--- Log opened Sun Jun 11 00:00:27 2023
00:34 < hprmbridge> kanzure> hallucinated libraries as attack vectors https://twitter.com/llm_sec/status/1667573374426701824
02:50 < alethkit> kanzure: It is all virtue ethics
02:51 < alethkit> also, nothing is surprising about LLMs so far, which is very disappointing
03:02 < TMA> the surprising thing about LLMs is that the generated contents which is not entirely hallucinated
03:02 < TMA> s/which//
03:33 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has joined #hplusroadmap
04:30 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has left #hplusroadmap []
04:31 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
05:13 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has left #hplusroadmap []
05:13 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
05:46 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0:7d6a:7b17:f05a:28ec] has joined #hplusroadmap
06:35 < kanzure> "Cryo-mechanical RAM content extraction against modern embedded systems" https://wootconference.org/papers/woot23-paper3.pdf
08:03 -!- test__ [~flooded@31.13.189.243] has joined #hplusroadmap
08:07 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 250 seconds]
09:31 < muurkha> jaisel: "motivation" is just a term for whatever internal processes in a human increase the probability of taking action
09:32 < muurkha> curiosity is sometimes one such process, but sometimes it's not
09:55 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
09:55 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
11:12 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
11:16 -!- test__ [~flooded@31.13.189.243] has quit [Ping timeout: 255 seconds]
11:52 -!- FourFire [~FourFire@141.51-175-99.customer.lyse.net] has joined #hplusroadmap
12:01 < kanzure> notes from the foresight institute whole brain emulation workshop earlier this month https://docs.google.com/document/d/1Mn9mTqpVNklJoRii_aUr3CxxgAriwV0u7FKwktTlasw/edit#
12:08 < hprmbridge> lachlan> Interesting
12:08 < hprmbridge> lachlan> You weren’t there, Bryan, right?
12:09 < kanzure> no
12:10 < kanzure> my notes would not be so short :)
12:10 < hprmbridge> lachlan> haha, thought so
12:10 < kanzure> they sent out the link in a public tweet
12:10 < hprmbridge> lachlan> I was in London at the same time as the meeting and considered trying to get in
12:11 < kanzure> just from a quick read i'm not sure there's any new ideas in there, maybe the "use algorithms to fill in lossy data from scans" is kinda new?
12:12 < kanzure> "Science doesn’t work: there’s always the explanation “the nanodust did it”. Like fake photos, except also for your direct experience of reality. No longer coherent physical laws."
12:13 < hprmbridge> lachlan> That would be unfortunate for human agency
12:14 < kanzure> "WBE may run into restrictions on compute due to restrictions on AGI"
12:14 < hprmbridge> lachlan> Well, it is AGI
12:15 < kanzure> demonstrating memory preservation, localization, extraction, and reconstitution in silico seems like the only thing that matters at this stage
12:16 < hprmbridge> lachlan> Is there any reason to expect memories to be a local feature and not require global emulation of an entire brain?
12:17 < kanzure> i could ask you the opposite
12:17 < hprmbridge> lachlan> You should
12:18 < kanzure> "In order to get ground truth data for WBE methods and set up a testing/validation pipeline, the “prettyprinter” will take an artificial neural network (ANN), convert it into a biologically plausible compartment model (e.g. Hudgin-Huxley), convert that into computer graphics, and then produce a simulated imaging stack as might come from real scanning methods. The challenge is then to use ...
12:18 < kanzure> ...imaging and reconstruction methods to reconstruct the compartment model or extract hidden information put into the ANN."
12:18 < kanzure> "We will make a computer program that will take something like ChatGPT2 and turn it into a 3D model with 3D neurons, 3D tubes connecting them, and cartoon synapses "globs" connecting the tubes together. We call this program a "prettyprinter" because it's like other programs that take things like tables of numbers and make text files that look more aesthetically pleasing. The prettyprinter will ...
12:18 < kanzure> ...also be able to create videos of the 3D network actually functioning. In the case of chatGPT2, imagine a big pile of ~100,000 3D semi-transparent neurons with around 1.7 billion 3D synapses that light up in different patterns depending on what input text chatGPT2 got. The prettyprinter can also produce simulated microscope slides that you could obtain by slicing the 3D network up and imaging ...
12:18 < kanzure> ...it. The prettyprinter is a good idea because you can use it to test methods of uploading with "password extraction" experiments. You train chatGPT2 to remember a specific password, then use the prettyprinter to spit out microscopy data and simulated activation data, then use your uploading method to see if you can get the password back. You can then make the whole process more challenging by ...
12:18 < kanzure> ...making the microscopy / activation data more "blurry", making the password more complicated, and using a different (either simpler or more complicated) network than chatGPT2."
12:19 < kanzure> hmm, that's kind of interesting, but i would be more interested in memory from a human brain or an animal instead
12:19 < hprmbridge> lachlan> I’m not really sure why that’d make sense
12:19 < hprmbridge> lachlan> A transformer doesn’t have 3d structure
12:19 < kanzure> it's certainly something you could do if you are willing to make certain assumptions about the equivalencies or isomorphism between modern ML and biological memory tissue
12:22 < hprmbridge> lachlan> It might be more valuable if the training process incentivized the networks to make “local” connections, for reasonable definition of “local”
12:23 < kanzure> for organoids or neural tissue cultures, what you could do if you had an unlimited budget is try to evolve a neural cell line that had some sort of interpretability (either through synapse connectivity imaging, or RNA sequencing, or methylation sequencing, etc), and just favor cell lines that produce biological neural networks that can upload some amount of data through that method.
12:23 < kanzure> it wouldn't work for existing biological human brains, but it would be at least something for future brains
12:23 < kanzure> if you could just reprogram brains to do whatever, one of the things oyu would probably want is for neurons to debug themselves and spit out their own connection graph as encoded in a sequence of motor neuron action potentials
12:27 < hprmbridge> lachlan> We probably will need to do that to determine a behavioral model neurons and non-neuron neuronal cells
12:27 < hprmbridge> kanzure> or co-train biological neural networks to only learn bits of data iff the other cultured cells can predict what the other one has learned based on connectivity; eg only interpretable connectivity for the data being learned will be learned by the first culture.
12:28 < hprmbridge> lachlan> WBEs probably look like small ANNs that govern the behavior of every cell in an emulated brain
12:39 < hprmbridge> kanzure> are there any papers going over the encoding of information solely from connectivity and how that could work even in theory?
12:39 < hprmbridge> kanzure> not accounting for synaptic strengths or weights
12:41 < hprmbridge> lachlan> I trained spiking neural networks for my undergrad thesis with synaptic weights that could be -1, 0, or 1
12:42 < hprmbridge> lachlan> That’s nearly connectivity-only training
12:43 < hprmbridge> lachlan> But they were liquid state machines, so there is external weighting now that I think about it
12:48 < hprmbridge> kanzure> you could extract information visually by synapse size, cell shape, cell diameter, connectivity. or you could have motor neurons embedded throughout and have it scrunch/de-scrunch like chromatophore control in octopus.
14:48 -!- dustinm [~dustinm@static.38.6.217.95.clients.your-server.de] has quit [Quit: Leaving]
15:00 -!- dustinm [~dustinm@static.38.6.217.95.clients.your-server.de] has joined #hplusroadmap
16:29 -!- stipa_ [~stipa@user/stipa] has joined #hplusroadmap
16:31 -!- stipa [~stipa@user/stipa] has quit [Ping timeout: 265 seconds]
16:31 -!- stipa_ is now known as stipa
18:39 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has quit [Ping timeout: 255 seconds]
18:56 < hprmbridge> jaisel> is anyone aware of any study/work type groups in seattle where people come together to discuss/work together for math/any hardware related fields
18:56 < hprmbridge> jaisel> ik its a broad question but am here for the summer and would love to meet and work with other motivated people in the area.
19:03 < hprmbridge> w. portr> if you go on lesswrong, they have a meetup list you can browse for your area.
19:04 < hprmbridge> w. portr> lesswrong.com/community
19:05 < hprmbridge> jaisel> cool! thx 🙂
19:53 -!- acertain [sid470584@2a03:5180:f:4::7:2e38] has quit [Server closed connection]
19:53 -!- acertain [sid470584@id-470584.hampstead.irccloud.com] has joined #hplusroadmap
20:03 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0:7d6a:7b17:f05a:28ec] has quit [Quit: Leaving]
20:30 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
20:30 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
23:30 < fenn> "Killed by Bad Philosophy"
23:30 < fenn> Sartre, an arrow through my heart
23:30 < fenn> (yes i know it's about deathism)
23:38 < hprmbridge> w. portr> sartre, camus and dawkins dragged my ass out of catholicism
23:49 -!- test__ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
23:53 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 255 seconds]
--- Log closed Mon Jun 12 00:00:28 2023