--- Log opened Sat Dec 10 00:00:41 2022
00:06 < maaku> eh there's still a human in the loop guiding the search
00:06 < maaku> but yeah, it's 90% of the way there
00:07 < maaku> I'm a bit dumbfounded by the people prognosticating about whether we'll have AGI in 2023, 2028, 2050, etc.
00:07 < maaku> we had AGI in 2021. possibly a few years earlier
00:08 < maaku> "oh but we're not all paperclips or living in utopia"
00:08 < maaku> well, maybe your idea about what AGI will mean is off
00:31 -!- fenn [~fenn@bryan.fairlystable.org] has quit [Changing host]
00:31 -!- fenn [~fenn@user/fenn] has joined #hplusroadmap
02:28 < TMA> maaku: which AGI we have had in 2021?
04:11 < maaku> Transformer architectures are capable of performing generalized problem solving and unrestricted knowledge transfer
04:11 < maaku> Which is the definition of AGI by the people who coined the term.
04:43 < maaku> Transformer is a completely domain independent unsupervised model learning architecture which is able to map learned models to new domains on the fly.
04:44 < maaku> Go back and read AGI conference notes, and you’ll see that this is exactly what people have been trying to figure out how to build for decades.
04:45 < maaku> Goal posts have moved and people think of AGI as some sort of skynet like program with agency, which is shifting goal posts
04:46 < maaku> But even so, adding agency is NOT the hard part. That’s a solved problem! As long as you don’t care about value alignment at least.
04:47 < maaku> Embodying a transformer network to have precepts and actions (e.g. hooking up GPT to the web) would be an intern level project.
04:48 < maaku> Giving it agency is literally 3 lines of Python.
04:50 < maaku> There would be some oddities around the fact that training transformer networks is done offline and in batch, with limited opportunity for online one-shot learning. So it wouldn’t think like a human, and it’s ability to learn would be high-latency. But it would work.
04:52 < maaku> Has anyone actually done this? Idk. But as a problem it has been solved since at least last year, and with the release of GPT-3 and DALL-E 2 we have had the hard parts done.
05:35 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has joined #hplusroadmap
05:36 < muurkha> the humans are also mostly trained offline and in batch
05:36 < muurkha> I mean their long-term learning rate during online tasks is not literally zero but it's damned close
05:39 < muurkha> fenn: I think it was in 02003 that he started working at Google, at which point he began to enthusiastically remain silent
05:41 < muurkha> I think it's maybe still being debated as to whether GPT-3 was actually performing generalized problem solving well enough to count
06:40 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0::a324] has joined #hplusroadmap
08:04 < kanzure> https://makepeoplebetterfilm.com/
08:07 < kanzure> "Sheehy is slightly more sympathetic, and even optimistic about what lies ahead... 'history will indeed show that ethics is on his side'."
08:13 < muurkha> .t http://bitterdb.agri.huji.ac.il/
08:13 < saxo> 403 Forbidden
08:13 < muurkha> that's too bad.  that was the database of bitter compounds
09:43 -!- codaraxis___ [~codaraxis@user/codaraxis] has joined #hplusroadmap
09:47 -!- codaraxis__ [~codaraxis@user/codaraxis] has quit [Ping timeout: 260 seconds]
10:18 -!- A_Dragon is now known as Festive_Dragon
10:27 < kanzure> .title https://nitter.nl/jd7h/status/1601535920875966464
10:27 < saxo> Judith van Stegeren (@jd7h): 'Georgi Gerganov has created a lightweight open source port of OpenAI's Whisper model that can run locally on your CPU. It's fast, fast enough to do real-time transcription of audio. I've tried it out on an old conference talk video from FDG 2021 and the results are impressive!'|nitter
10:27 < kanzure> https://github.com/ggerganov/whisper.cpp
10:34 < kanzure> .title https://nitter.nl/EleanorSheekey/status/1601592574875049984
10:34 < saxo> Eleanor Sheekey (@EleanorSheekey): 'Does memory have a mechanical basis?Here I chat with Prof @bengoult about his MeshCODE theory that considers the protein Talin as an information storing molecule. We also chat ECM stiffness, aging and memory loss.--> https://www.youtube.com/watch?v=QS8OXkcOb0U #scicomm #AcademicYouTube'|nitter
10:34 < Muaddib> [QS8OXkcOb0U] Memory molecules, Neural Computation & the ECM - Prof Ben Goult (66:25)
11:37 < kanzure> "Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models" https://www.biorxiv.org/content/10.1101/2022.12.09.519842v1
11:45 -!- codaraxis__ [~codaraxis@user/codaraxis] has joined #hplusroadmap
11:49 -!- codaraxis___ [~codaraxis@user/codaraxis] has quit [Ping timeout: 256 seconds]
12:19 -!- codaraxis___ [~codaraxis@user/codaraxis] has joined #hplusroadmap
12:23 -!- codaraxis__ [~codaraxis@user/codaraxis] has quit [Ping timeout: 260 seconds]
12:41 < fenn> whisper on raspberry pi? eesh
12:41 < fenn> that was quick
12:47 < kanzure> lopp on twitter tested it against my transcript of roasbeef, i still think that is an unfair comparison
13:09 -!- luna_ [~luna@2a01:4c8:a3:acdf:6841:861b:87ff:dfe5] has joined #hplusroadmap
13:09 -!- luna_ [~luna@2a01:4c8:a3:acdf:6841:861b:87ff:dfe5] has quit [Changing host]
13:09 -!- luna_ [~luna@user/luna/x-4729771] has joined #hplusroadmap
13:30 -!- srk [~sorki@user/srk] has quit [Ping timeout: 252 seconds]
13:37 < fenn> i was able to compile whisper.cpp without AVX. it works, but it's 16x slower than realtime with the smallest model
13:41 < fenn> on a relatively old laptop
13:53 < fenn> hm throwing gobs of cores at it actually slows it down
13:58 < muurkha> interesting
13:59 -!- srk [~sorki@user/srk] has joined #hplusroadmap
13:59 < fenn> nevermind i was probably causing some kind of resource contention by maxing out the system
14:01 < fenn> still, on a 32 core server the best i can do is 1/3.85x realtime (without AVX)
14:03 < fenn> strangely, it gives different text output on subsequent runs. on the JFK sample, sometimes it splits the line after "ask not what your country can do for you," and sometimes it doesn't
14:05 < fenn> [00:00:00.000 --> 00:00:11.000]   And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
14:06 < fenn> [00:00:00.000 --> 00:00:05.500]   And so my fellow Americans ask not what...
14:06 < fenn> [00:00:05.500 --> 00:00:12.380]   your country can do for you. Ask what you can do for your country.
14:09 -!- Molly_Lucy [~Molly_Luc@user/Molly-Lucy/x-8688804] has quit [Quit: Textual IRC Client: www.textualapp.com]
15:04 -!- cc0 [~cc0@2a01:4f9:c010:cf0b::1] has quit [Remote host closed the connection]
16:07 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has quit [Ping timeout: 260 seconds]
16:32 -!- codaraxis__ [~codaraxis@user/codaraxis] has joined #hplusroadmap
16:35 < maaku> fenn: does it have other vectorization options, or is it just AVX or nothing?
16:36 -!- codaraxis___ [~codaraxis@user/codaraxis] has quit [Ping timeout: 260 seconds]
16:37 -!- codaraxis [~codaraxis@user/codaraxis] has joined #hplusroadmap
16:40 -!- codaraxis__ [~codaraxis@user/codaraxis] has quit [Ping timeout: 256 seconds]
17:20 < kanzure> "Reversal of epigenetic aging and immunosenescent trends in humans" (fahy thymus regeneration article from 2019) https://onlinelibrary.wiley.com/doi/full/10.1111/acel.13028
18:11 < fenn> maaku: ggml.c has the following ifdefs: __HAIKU__ GGML_USE_ACCELERATE __ARM_NEON __wasm_simd128__ __F16C__ GGML_PERF GGML_GELU_FP16 __APPLE__
18:12 < fenn> at least some of those are other vectorization architectures, but i'm not really familiar with the available options in the first place
18:12 < fenn> nothing else on intel anyway
18:13 < fenn> uh, f16c is some kinda conversion function to make it work with AVX
18:15 < fenn> i get the feeling AVX is supposed to work, and it's some kind of linux bug that they haven't gotten around to fixing
18:17 < maaku> yeah those seem like arm extensions
18:28 < luna_> What's the rationale for the theory of longevity escape velocity? I mean yeah, there could be a breakthrough that raises global LE by more than a year and takes less than a year but how does that imply that breakthroughs of similar magnitude will keep happening indefinitely? Doesn't it depend on human ingenuity?
19:02 -!- Mabel [~Malvolio@idlerpg/player/Malvolio] has quit [Ping timeout: 256 seconds]
19:02 < jrayhawk> all breakthroughs enable further breakthroughs. for instance, increasingly sophisticated versions of CRISPR and gene synthesis enable essentially *everything*. Additionally, biological control also involves control of cognition itself.
19:03 < jrayhawk> that said, setbacks can occur. there are periods of biology where we forgot how vitamin c works, or cod liver oil.
19:03 < kanzure> "longevity escape velocity" is more useful as an introductory concept for people who don't understand all the forms that immortality could take
19:04 < kanzure> also, i've advocated in the past that the most useful genetic longevity enhancements are most likely to only accrue to germline (meaning, children and not you)
19:04 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0::a324] has quit [Quit: Leaving]
19:08 < jrayhawk> the question at that point becomes "can in-situ modification catch up to the capabilities of germline modification in the time it takes to grow a new adult (say, 15-20 years)?"
19:08 < nsh> (it's a lot more efficient to just get over the delusion that you're mortal to begin with)
19:09 < nsh> (doesn't really require any advanced praxic technology at all; can be achieved in a cave in a matter of years depending on aptitude)
19:09 < luna_> Interesting. How? Why isn't there proof that anyone has ever done it?
19:10 < kanzure> don't mind nsh he just rambles like that
19:11 < kanzure> we only let him out at night
19:12 < nsh> there's as much proof as there is pudding :)
19:12 < luna_> I imagine. I'm not convinced that once improvement of life extension technology reaches a certain rate it implies that it can be sustained at that rate. I hear a lot of people talk about LEV as immortality, you mention it's just an introductory concept, but I'm missing how it's valid at all. I get that breakthroughs build on previous ones but how does that guarantee a sustained rate of progress?
19:13 < nsh> if a child refuses to believe that numbers can be added together and isn't interesting in giving it a go, is that end of arithmetic?
19:13 < kanzure> LEV should be considered a post-factual explanation: you are looking at the future and you see someone who was born today, but he's alive 300 years from now. how did that happen? LEV is an explanation for that.
19:13 < nsh> one thinks not
19:14 < kanzure> luna_: i don't think anyone is saying that they know how to achieve "longevity escape velocity", only that there in theory could be a sequence of developments that could enable the first billionyearaire to have already been born
19:15 < luna_> Well there are current predictions of LEV by 2030
19:15 < luna_> It sounds preposterous to me.
19:15 < kanzure> wait do you have a positive opinion of most other future predictions??
19:16 < luna_> Not usually.
19:16 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
19:16 < kanzure> futurists in their role as the great prognasticators have not exactly positioned themselves for accuracy nor to financially benefit or anything from their predictions
19:17 < kanzure> anyone can sit around and draw an exponential on a chart
19:18 < kanzure> at this point i'm just rambling, nevermind
19:18 < kanzure> luna_: https://diyhpl.us/wiki/genetic-modifications/ (under the 'longevity' section)
19:18 < luna_> I'm not the best judge of it. I was wondering if despite how it intuitively appears to me, there might be something to it.
19:22 < kanzure> i don't think aubrey said it was a guaranteed trajectory or anything
19:23 < kanzure> velocity here is hard to measure too
19:24 < kanzure> the wikipedia article is a little odd: "At present, more than one year of research is required for each additional year of expected life"... i mean, maybe, by like a factor of several 100,000 researcher life-years each year, according to global annual QALY increases? certainly not a mere "more than one year of research".
19:25 -!- cc0 [~cc0@2a01:4f9:c010:cf0b::1] has joined #hplusroadmap
19:31 < luna_> I interpret the article as referring to the rate of progress with the current level of resource allocation taken into account.
19:32 -!- cc0 [~cc0@2a01:4f9:c010:cf0b::1] has quit [Remote host closed the connection]
19:40 < nsh> (as to the how, an expeditious means is to receive the pointing-out instruction of dzogchen)
19:52 < fenn> luna_: prediction markets tend to give more reliable answers than "futurists" although markets can be way off as well, especially with speculative technology
19:54 < fenn> this market gives 50/50 odds on *any* useful longevity technology by 2030, not even reaching 10 years of life extension https://www.metaculus.com/questions/3458/will-there-be-a-culturally-significant-development-in-aging-research-by-2030/
19:54 < fenn> the quote there from a specialist in the field is even more pessimistic
19:54 < fenn> so i wonder who you're listening to that is so optimistic
19:55 < luna_> A lot of people quote Kurzweil on LEV. He originated the 2030 prediction in 2018.
19:55 < fenn> unfortunately for the 'AI will solve everything' crowd, actually testing hypotheses about aging in large animals will take a long time, simply because it takes a long time to get a long enough baseline that your imprecise measurements mean anything
19:57 < fenn> even if AI somehow comes up with a general theory of aging right now, we won't know if it actually works in humans by 2030, barring some crazy spontaneous phenotype rejuvenation effect
19:58 < fenn> like, how do you know you haven't just activated dormant stem cells
19:58 < fenn> and then you run out of stem cells and die
19:59 < fenn> there are actually a lot of promising theories kicking around, but they haven't really been tested
19:59 < fenn> a big part of the problem is you don't know whether a given intervention is affecting the root cause, or a downstream causal node
20:00 < fenn> you could get enormous benefits, and still not have figured out the true root cause of aging, which of course means you'll die in the end
20:01 < fenn> so "longevity escape velocity" is the bet that we will get enough of those type of interventions that we stay alive long enough to find the true root cause of aging (and fix it)
20:01 < luna_> Yeah apparently the singularity is coming around 2030 as well
20:02 < fenn> and really it's a bet with no downside. you could get hit by a bus tomorrow, so live your life as if it mattered today.
20:02 -!- cc0 [~cc0@2a01:4f9:c010:cf0b::1] has joined #hplusroadmap
20:02 < luna_> I view transhumanism with at least my baseline level of skepticism
20:02 < luna_> Interesting as fuck though
20:02 < luna_> Conspiracy theories about it are hilarious as well.
20:03 < fenn> saw this one yesterday https://www.youtube.com/watch?v=1Qzc8Mr3xb4
20:03 < Muaddib> [1Qzc8Mr3xb4] Mgs 2 predicted this 20 years ago (Selection for Societal Sanity) (7:15)
20:03 < fenn> MGS = metal gear solid, the video game
20:04 < fenn> throwing in "oh, we solved uploading, and have been regularly genetically engineering humans for the past 50 years" casually
20:06 < fenn> the conspiracy theories are because a few transhumanists got lucky in the web 2.0 tech bubble
20:06 < luna_> Lol
20:07 < luna_> Throwing in the covid vaccine
20:07 < fenn> now there's this perception of "transhumanist elites" which is pretty far from the case, but it has a sexy appeal so the meme spreads
20:07 < luna_> Ah makes sense
20:09 < fenn> you wouldn't want to be ruled and enslaved by a clueless incompetent regressive fuckwit with absolutely no vision. in contrast, a supervillain genius looks positively attractive
20:09 < fenn> well, the rhetorical "you"
20:10 < fenn> i wonder what's taking vat-grown organs so long
20:25 < fenn> https://www.youtube.com/watch?v=GqJYMgAcc0Q
20:25 < Muaddib> [GqJYMgAcc0Q] Bioengineers clear major hurdle on path to 3D printing replacement organs (2:46)
20:26 < fenn> so there's this 3D printing approach, which is great. but we also have the recellularization of animal organs route, which is a lot more straightforward
20:27 < fenn> that whole sub-field kinda dropped off the radar in 2014 or so
20:33 < superkuh> Now it's for growing animal-less meat.
20:33 < fenn> that's not even what i'm talking about though
20:33 < superkuh> Oh. Should've read back more. Disr... right.
20:33 < fenn> organoids are a third approach. i'm talking about reusing the collagen scaffold from a pig organ, but seeded with human progenitor cells
20:34 < fenn> "Pigs have anatomies strikingly similar to that of humans. Pigs are generally healthier than most primates and they're extremely easy to breed, producing a whole litter of piglets at a time. Moral objections to killing pigs are fewer since they're slaughtered for food."
20:35 < fenn> pig organs are basically good enough already. there are even pigs that have been genetically engineered to remove endogenous retroviruses, although that shouldn't matter if you're decellularizing
20:35 < fenn> so what's the hold up?
20:35 < fenn> you'll have to figure out how to seed a scaffold anyway to get 3d printing based approaches to work
20:39 < fenn> you might need to incubate the organs in a pig until they're ready to be implanted. this could be done with brain-dead humans who will otherwise be taken off life support, but that approach has legal and ethical complications
20:41 < fenn> i'm over here imagining a universal organ donor tissue type, when we don't even have artificial blood yet
20:45 < fenn> "many animals, such as pigs, have a shorter lifespan than humans, meaning that their tissues age at a quicker rate. (Pigs have a maximum life span of about 27 years."
20:49 < fenn> "In September 2021, surgeons led by Robert Montgomery performed the first genetically engineered pig kidney xenotransplant to a brain-dead human at NYU Langone Health with no sign of immediate rejection (partly because the pig thymus gland was transplanted as well). The kidney was procured from a pig with only a single gene modification: the removal of alpha-gal."
21:01 < fenn> heh robert montgomery himself has received a transplanted heart. i wonder how much "ethical review" was required to make that happen
22:13 -!- berndj [~berndj@197.189.254.139] has quit [Quit: ZNC - http://znc.in]
22:14 -!- berndj [~berndj@197.189.254.139] has joined #hplusroadmap
22:47 < maaku> medical advances tend to come in punctuated equilibria. the creation of antibiotics was a step function. i suspect RNA vaccines (when applied to cancer, flu, etc.) will be too
22:48 < maaku> LEV is a great way to get across the idea that it is possible we might live forever
22:48 < maaku> to someone not already accostomed to the idea
22:48 < maaku> but I don't think it is very descriptive of reality
23:37 -!- Guest58 [~Guest58@115.186.158.117] has joined #hplusroadmap
23:38 -!- Guest58 [~Guest58@115.186.158.117] has quit [Client Quit]
--- Log closed Sun Dec 11 00:00:42 2022