--- Log opened Sat Jul 15 00:00:59 2023
01:11 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has joined #hplusroadmap
01:53 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0:ad35:f6ea:8157:cbf9] has joined #hplusroadmap
04:14 -!- stipa [~stipa@user/stipa] has joined #hplusroadmap
04:20 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has left #hplusroadmap []
04:23 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
04:49 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has left #hplusroadmap []
06:43 < docl> why does everyone remember the fanfic instead of the rationality sequences
06:44 < docl> like, people who think they are not useful must all not have read them or something
06:49 -!- L29Ah [~L29Ah@wikipedia/L29Ah] has joined #hplusroadmap
08:13 < hprmbridge> kanzure> https://www.npr.org/sections/health-shots/2023/07/15/1184298351/conception-human-eggs-ivg-ivf-infertility
08:15 < L29Ah> i wonder why english-language sites love to write articles full of "On a cloudy day on a gritty side street near the shore of San Francisco Bay, a young man answers the door at a low concrete building." bullshit
08:16 < mrdata> dont you want a story to give you a picture of the scene?
08:17 < L29Ah> no, i want to read about the "Startup aims to make lab-grown human eggs, transforming options for creating families"
08:18 < mrdata> but it's NPR
08:18 < mrdata> not Healthline
08:19 < juri_> I once had an NPR story about what i used to do. now i stare at math all day.
08:44 -!- test_ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
08:47 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 246 seconds]
09:06 < hprmbridge> kanzure> https://www.sam-rodriques.com/post/optical-microscopy-provides-a-path-to-a-10m-mouse-brain-connectome-if-it-eliminates-proofreading
09:09 < hprmbridge> kanzure> a tweet thread about yamanaka factors and tumors https://twitter.com/CharlesMBrenner/status/1680245710560821248
09:42 < hprmbridge> alonzoc> The journalists need to flex their creative writing degrees somehow
09:46 < hprmbridge> alonzoc> The sequences are useful, well some of them at least, the issue I have is most of even the best sequences aren't really saying anything new. Lesswrong has a habit of reinventing already existing philosophy under a new name. However to be honest they do cut out a lot of the obscuritory BS.
09:46 < hprmbridge> alonzoc> But many of the sequences are just as bad as self-important philosophy books which coat fairly simple insights in paragraphs to make them seem deep
10:33 < docl> well, much as I can see the self importance criticism, that cutting through obscuratory BS on existing philosophy is basically what I think of when people say EY hasn't done anything useful. it seemed useful to me when I was reading it, anyway. also, he has done a lot of moving the overton window on AI safety. I'm not saying he's right, but people are putting more work into proving he isn't. I suspect 
10:33 < docl> this actually accelerates progess because it means the fears can be addressed on a more explicit level. ironically, sam altman credits EY with getting him interested in AI (much to EY's apparent chagrin)
10:41 < hprmbridge> alonzoc> Ehh he's moved the overton window amongst some people but I think he's done a bit more harm than good. Otherwise intelligent people dismiss AI safety because it's associated with the lesswrongers whole beleif system. The AI safety comes with baggage from stuff like acausal trades, longtermism, rationalism etc. The lesswrongers spent their "weirdness points" a bit to hard.
10:41 < hprmbridge> alonzoc> They've helped but they've also harmed.
10:41 < hprmbridge> alonzoc>
10:41 < hprmbridge> alonzoc> Anyways I do like some of the work that's on lesswrong I just have issues with a lot of the safety people. I think I agree heavily with the transcript about e/acc on the diyhpl.us sans the slight elon fanboyism
10:42 < hprmbridge> alonzoc> But like the works of people like Garrabrant really shouldn't be ignored just because you've got issues with Yud or some of lesswrongs sociology
10:45 < docl> who's garrabrant?
10:45 < docl> oh, think I found him http://scott.garrabrant.com/
10:46 < docl> anything of his you particularly recommend?
10:57 < hprmbridge> alonzoc> His paper on logical induction is a must read if a bit of a slog because it insists on imo an annoying metaphor of a betting market, another good paper in the vein of logical induction is https://arxiv.org/abs/1904.11099 .
10:57 < hprmbridge> alonzoc>
10:57 < hprmbridge> alonzoc> Another interesting bit of research is his work on causality through an alternative formalisation https://www.lesswrong.com/s/2A7rrZ4ySx6R8mfoT which imo really nicely disentangles the differences between different decision theories even though I haven't completely grokd everything he's written as of yet
11:05 < docl> nice, will take a look at those
11:11 < docl> haven't read the paper yet, but betting markets vaguely annoy me. any thoughts on why it's an annoying metaphor in this kind of context?
11:14 < hprmbridge> alonzoc> I think it obscures the algorithm, it makes it overly complex. Other brute force "reified ideals" for a concept like levin search don't have anything more than necessary.
11:14 < hprmbridge> alonzoc>
11:14 < hprmbridge> alonzoc> Like markets do aproximate bayesian information aggregation as a dynamical process with eventual convergence, but they're one implementation of such a dynamical system. It's a case of all markets are inductors but not all inductors are markets
11:15 < hprmbridge> alonzoc> Like any inductor can be implemented as a market but it's kinda a contrived transformation imo
11:17 < hprmbridge> alonzoc> Also real markets have a bunch of behaviours which are far from ideal they only converge in the limit. Traditional markets with noise traders exhibit flocking behaviour which creates bubbles and crashes even if the market is arbitrage free
11:18 < hprmbridge> alonzoc> I really need to fix my irc client
11:25 < docl> levin's search http://www.scholarpedia.org/article/Universal_search
11:25 < docl> I see what you're saying. makes sense.
11:26 < hprmbridge> alonzoc> Oh schmidhubers oops papers are already worth a read and his C code is the worst I've ever seen
11:26 < docl> lol OK
11:27 < docl> is this stuff what ends up used in LLM designs? or is that a different direction entirely?
11:29 < hprmbridge> alonzoc> Completely different direction, a lot of it is pre deep learning. But any AI system fundementally approximates solomonoff induction so they're useful at the very least for theoretical analysis, and at some point we're gonna have to try learn program like structures to take neural nets further I suspect and then there's some real issues with naïve gradient descent
11:33 < hprmbridge> alonzoc> https://www.youtube.com/watch?v=IW4LjjAWrO4 this is worth a watch. They ended up deriving learning rules or atleast error prop rules for Turing machines from first principles. It works for small synthesis problems but runs into issues with an explosion of minima. In a later paper "geometry of program synthesis" they use algebraic geometry and singular learning theory to talk about the search space
11:33 < hprmbridge> alonzoc> a bit more and it's suspected the issue is to do with the structure of the parameter spaces singularities which might explain why normal regularisers based on description length don't help gradient descent in program space. I also personally suspect there's some links to phase transitions in SAT problems
11:53 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
11:57 -!- test_ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 252 seconds]
13:28 < muurkha> what's the current status of neural heuristics for accelerating SAT solvers?
13:30 < hprmbridge> alonzoc> Nothing particularly good imo, but if you're interested in something that's "neural" by a loose defenition of learned circuits you might be interested in knowledge compilation
13:30 < hprmbridge> alonzoc> A good talk on this is: https://youtu.be/3PrYYLppjXA
13:30 < Muaddib> [3PrYYLppjXA] Three Modern Roles for Logic in AI | Adnan Darwiche | PODS 2020 (86:27)
13:31 < hprmbridge> alonzoc> Basically you can compile sat problems to circuits which are compressed sets of solutions where you can enumerate solutions in polynomial time and do other operations on them. These are used quite a lot in stuff like probabilistic prolog etc, it's also useful for caching common subproblems
13:31 < muurkha> thanks!
13:32 < hprmbridge> alonzoc> https://youtu.be/8JI-KTJKr3w
13:32 < Muaddib> [8JI-KTJKr3w] Approximate Knowledge Compilation by Online Collapsed Importance Sampling  | NeurIPS 2018 (13:28)
13:32 < hprmbridge> alonzoc> This is also a good talk
13:39 < muurkha> maybe there's a paper I can read since it's from neurips
13:39 < hprmbridge> alonzoc> You can read any paper if you try hard enough
13:40 < hprmbridge> alonzoc> Lol, I'm pretty sure the big LLMs are being trained off libgen
13:41 < muurkha> I mean most youtube videos don't have an associated paper
13:42 < hprmbridge> alonzoc> Tru, the one's I've shared all do iirc
13:49 < hprmbridge> alonzoc> The differential linear logic paper takes a bit of finding but can be found here https://arxiv.org/abs/1805.11813 there are a few papers on this but they're all close together in the citation graph
14:09 < juri_> now that's the #hplusroadmap i've come to know and love. i feel like an idiot reading the backlog, as it should be. :)
14:14 < hprmbridge> alonzoc> Hmm I wonder how many tokens worth of conversational text data is in hplus' irc logs
14:15 < hprmbridge> kanzure> semantic search over the logs would be really nice.
14:16 < hprmbridge> alonzoc> Pretty useful probably, I've already been thinking about feeding all the logs from my chats with friends into a big semantic memory store. I've had quite a few discussions about LLMs with long term external memories
14:17 < hprmbridge> alonzoc> Tbh it should also reduce the needed LLM size as Wikipedia and shit could be stored in the "fast weights" of a neural RAM instead of being burned into the main models synapses
14:18 < hprmbridge> alonzoc> Like think about the relative information content of "all the internets knowledge" and "the schema and rules of English"
16:40 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
16:41 < hprmbridge> alonzoc> I mentioned schmidhubers oops algorithm earlier and said how bad the c was. Thought I'd post a link
16:41 < hprmbridge> alonzoc> https://people.idsia.ch/~juergen/oopscode.c
16:41 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
16:41 < muurkha>         define_bytecode_niladico ret, 2
16:42 < muurkha> oops
16:42 < muurkha> /** This code layout is  an instance of "crystalline formatting". **/
16:42 < muurkha> /**** J. Schmidhuber's philosophy behind crystalline formatting is:*/
16:42 < muurkha> /** Some programmers prefer to debug code that looks nice. *********/
16:42 < hprmbridge> alonzoc> Yes... Crystalline formatting.
16:45 < muurkha> it doesn't seem like bad code, just eccentric
16:46 < muurkha> the amount of effort expended in writing this comment block is staggering:
16:46 < muurkha> /*we will now create initial perhaps recursive codes that go to q[]*/
16:46 < muurkha> /*enter(A) to assign code & Qnum and 0 start-address to Q's name, A*/
16:46 < hprmbridge> alonzoc> Tbh that's a better description, the c preprocessor macro overuse kinda annoys but it's.. defenitely a unique piece of code.
16:46 < hprmbridge> alonzoc> I do appreciate the amount of commenting effort tho
16:46 < muurkha> /*enter(A) to assign code & Qnum and 0 start-address to Q's name, A*/
16:47 < muurkha> /*enterv and enterw: enter code such as in5(),x34(), cp3() & C4C2()*/
16:47 < muurkha> /*ENTER (foo) states that foo is one of the codes in optimal search*/
16:47 < muurkha> /*decl(m,n,foo,body) makes footext & fooaddress; foo code: m inputs*/
16:47 < muurkha> /*n outputs; at most 1 new name in a body, for 1fold crossrecursion*/
16:48 < muurkha> if you're using an IRC client with a prooortional font you're missing the whitespace poetry
16:48 < muurkha> thoough arguably "in5(),x34()" and "1fold" is cheating
16:49 < hprmbridge> alonzoc> My IRC config is broke so I'm using discord for now, need to patch my the SASL stuff
16:50 < hprmbridge> alonzoc> Emacs erc mode had bitrotted in my config
16:51 < hprmbridge> alonzoc> But yeah oops.c is a unique piece not many appreciate or even know about. A particularly interesting algorithm implemented in a unique coding style
16:52 < hprmbridge> alonzoc> Still a pain to read through, to fully understand it I basically ended up transpiling it manually
16:53 < muurkha> it's kind of like an unfamiliar dialect of Lisp
16:54 < muurkha> it sort of reminds me of procmail too
16:55 < hprmbridge> alonzoc> I think it would be quite nice implemented in scheme, especially if you switch out the FORTH interpreter for a metacircular scheme interpreter
16:59 < muurkha> sounds interesting
17:16 -!- codaraxis [~codaraxis@user/codaraxis] has joined #hplusroadmap
17:40 -!- yashgaroth [~ffffffff@2601:5c4:c780:6aa0:ad35:f6ea:8157:cbf9] has quit [Quit: Leaving]
18:07 < hprmbridge> kanzure> " [T]he scary vision of the future isn't one where dinosaurs escape Isla Nubar and fly to the mainland, putting a healthy planet at risk, but instead a future where there aren't enough animals left to support food webs and ecosystems. And that includes humans, too... [I]t is our belief that it is possible to safeguard against or even stop that fatalist future vision using a similar approach in the
18:07 < hprmbridge> kanzure> original movie with some slight variations. It all goes back to genetics and a lot of what I learned about when I first met George..."
18:07 < hprmbridge> kanzure> this is how you get mammoths
18:11  * L29Ah is sad about the lack of attention/funding of Biosphere 2 and related projects
18:11 < L29Ah> mr Musk's Mars habitation simulator in desert/antarctic reality show when?
18:31 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
18:31 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
18:35 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has quit [Ping timeout: 264 seconds]
20:35 -!- codaraxis [~codaraxis@user/codaraxis] has quit [Quit: Leaving]
23:15 -!- helleshin [~talinck@108-225-123-172.lightspeed.cntmoh.sbcglobal.net] has joined #hplusroadmap
--- Log closed Sun Jul 16 00:00:00 2023