--- Log opened Sun May 28 00:00:13 2023
00:01 < jrayhawk> ##hplusroadmap/2014-11-10.log.xz:01:04 < fenn> oh jeez now you're saying lamarckian evolution is true?
00:01 < fenn> there are also gene cassettes, viral incorporation of mRNA, bacterial conjugation, probably other mechanisms. these all accelerate the spread of genes in active use more than silent genes
00:10 < fenn> obviously groups are subject to the consequences of actions of their members
00:10 < jrayhawk> ?
00:11 < fenn> uh nevermind, probably shouldn't drag a 9 year old argument into the present
00:26 < fenn> in conclusion, i know slightly more about epigenetics now than ten years ago. there is still a lot to learn
03:20 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has joined #hplusroadmap
03:30 -!- flooded [~flooded@146-70-115-163.pool.ovpn.com] has joined #hplusroadmap
03:34 -!- test_ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 265 seconds]
05:53 < fenn> https://neosensory.com/science/ seems like a bad decision to put it on the wrist but better than nothing
05:53 < fenn> "people who are deaf or hard of hearing can learn to identify sounds that are algorithmically translated into spatiotemporal patterns of vibration on the skin"
05:54 < fenn> they also discuss using it for artificial senses like infrared vision or abstract data streams
05:56 < fenn> the new iphones have a similar haptic device and cost less while doing more
06:03 < L29Ah> fenn: what haptic device? googling only suggests some vibration response settings akin to the stuff that is present in cellphones for over 30 years
06:04 < fenn> https://www.ifixit.com/News/16768/apple-taptic-engine-haptic-feedback
06:05 < fenn> this looks very close to what's shown in the neosensory buzz device https://www.precisionmicrodrives.com/wp-content/uploads/2021/10/lra-linear-vibrator-construction.original.jpg
06:06 < L29Ah> looks like an usual coin-type vibration motor that is used in smartphones for ~10 years
06:06 < L29Ah> what's so special about iphone's?
06:06 < L29Ah> the ifixit link lacks any photos of iphone's
06:07 < fenn> the article explains that it's bigger, and they tune the resnonant frequency of the oscillating voice coil mass to the resonant frequency of the phone body
06:08 < fenn> it's not a motor
06:08 < fenn> https://i0.wp.com/neosensory.com/wp-content/uploads/2021/09/8.22-Assembly-1.png looks similar
06:09 < L29Ah> https://www.precisionmicrodrives.com/wp-content/uploads/2021/06/coin-shaftless-vibration-motor-exploded-view.original.jpg oh indeed those are different
06:09 < fenn> that is a motor
06:10 < fenn> admittedly i can't actually tell from the cad rendering what's inside
06:14 < L29Ah> https://commons.wikimedia.org/wiki/File:IPhone_6s_-_Taptic_Engine_-_opened-93373.jpg
06:17 < L29Ah> https://upload.wikimedia.org/wikipedia/ar/f/f1/Taptic_Engines_sizes.jpg
06:21 < fenn> it's not like apple invented the concept, they're just the only ones to actually put it in their phones
06:22 < fenn> the reason i brought it up is the neosensory buzz wristband costs $1k
06:24 < fenn> each of those actuators is less than $10, so it seems like a bit much for "algorthms"
06:24 < fenn> the less algorithms the better in my opinion
06:24 < fenn> let the brain do the processing
06:25 < fenn> a lot of accessibility tech seems to focus on new users rather than peak performance, and that's a shame
06:26 < fenn> then you're on training wheels for the rest of your life
06:36 < jrayhawk> the other $990 is presumably for bureaucratic hoop-jumping for the FDA and FCC
06:38 < L29Ah> or they have a really small scale manufacturing
09:37 -!- juri__ [~juri@79.140.121.73] has joined #hplusroadmap
09:40 -!- juri_ [~juri@84-19-175-187.pool.ovpn.com] has quit [Ping timeout: 256 seconds]
09:41 -!- juri__ [~juri@79.140.121.73] has quit [Read error: Connection reset by peer]
09:41 -!- pharonix71 [~pharonix7@user/pharonix71] has quit [Ping timeout: 240 seconds]
09:42 -!- juri_ [~juri@79.140.121.73] has joined #hplusroadmap
09:43 -!- juri_ [~juri@79.140.121.73] has quit [Read error: Connection reset by peer]
09:43 -!- pharonix71 [~pharonix7@user/pharonix71] has joined #hplusroadmap
09:47 -!- juri_ [~juri@84-19-175-187.pool.ovpn.com] has joined #hplusroadmap
09:48 -!- test_ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
09:52 -!- flooded [~flooded@146-70-115-163.pool.ovpn.com] has quit [Ping timeout: 240 seconds]
10:06 < docl> https://www.alignmentforum.org/tag/mesa-optimization
10:33 < docl> huh, chatgpt is as about confused as me about the mesa- prefix as I am. since it's the spanish word for table. I guess you can picture the optimization happening on a surface, akin to the maxwell plaster cast of the thermodynamic surface
10:37 < docl> I tend to use gpt3.5 by default, might have to break out gpt4 for deeper insights... ah, it's suggesting mesa as similar to the geological term, something that rises above the landscape
10:39 < docl> OK, so human morals, reasoning, and so on are a mesa-optimizer where human evolution is the meta-optimizer
10:41 < docl> or maybe evolution is just a plain old optimizer, morals in general are a meta-optimizer, and specific morals / clusters of moral opinions in our society/societies are mesa-optimizer(s)
10:52 < docl> I have been exposing my daughter to numberblocks and other educational shows and teaching her counting concepts, since infancy really. she says some of the numbers out loud, although I'm guessing actually learning to coherently express math concepts like counting and addition is going to take a little longer. not optimizing purely for AI/AI safety here, there are lots of ways she can use math for life 
10:52 < docl> success.
11:08 < docl> Here's a brief convo with gpt4 that seemed useful for clarifying wrt mesa-optimizers. https://chat.openai.com/share/09f6ba76-426d-462f-87d5-d3c94fb4cbfb
11:32 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
11:32 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
11:36 < docl> worth noting the 2d landscape (north-south / east-west) is a simplification and you can have arbitrary number of dimensions (weights/parameters), with the verticality being the loss function. so rather than (x,y,z) it's more like (w1,w2,...,L). so basically the mesa is an elevated (higher than the minimum amount of loss) chunk of conceptual hyperspace, not strictly planar
11:40 < docl> (sorry to belabor the obvious, I just figure there's probably someone in the chat who finds that sort of thing to be a stumbling block. good old school LW about this: https://www.lesswrong.com/posts/2TPph4EGZ6trEbtku/explainers-shoot-high-aim-low)
11:44 < docl> BTW sharable chatgpt convos are a thing now
13:22 -!- A_Dragon [A_D@libera/staff/dragon] has quit [Killed (Stx (*Kiss of death*))]
13:22 -!- A_Dragon [A_D@libera/staff/dragon] has joined #hplusroadmap
13:50 < hprmbridge> nmz787> https://www.science.org/content/article/inspired-sea-and-sky-biologist-invents-new-kind-microscope
13:50 < hprmbridge> nmz787> https://www.nature.com/articles/s41587-023-01717-8
13:50 < hprmbridge> nmz787>  https://cdn.discordapp.com/attachments/1064664282450628710/1112483004795850812/41587_2023_1717_Fig1_HTML.webp
13:50 < hprmbridge> nmz787>  https://cdn.discordapp.com/attachments/1064664282450628710/1112483012844724244/41587_2023_1717_Fig2_HTML.webp
13:50 < hprmbridge> nmz787>  https://cdn.discordapp.com/attachments/1064664282450628710/1112483026874662962/41587_2023_1717_Fig3_HTML.webp
13:50 < hprmbridge> nmz787> "Inspired by the sea and the sky, a biologist invents a new kind of microscope
13:50 < hprmbridge> nmz787> The device can achieve clear images using samples suspended in any kind of liquid"
13:51 < hprmbridge> nmz787> "Reflective multi-immersion microscope objectives inspired by the Schmidt telescope"
13:54 < hprmbridge> nmz787> An hour of progress from yesterday and today (and Amazon overnight delivery)... My hydraulic quick connects are covered now https://cdn.discordapp.com/attachments/1064664282450628710/1112483994538360842/PXL_20230528_205239524.jpg
13:54 < hprmbridge> nmz787>  https://cdn.discordapp.com/attachments/1064664282450628710/1112484024716378112/PXL_20230528_205255453.MP.jpg
13:55 < hprmbridge> nmz787> Choosing those rubber things took way too long, why these people don't all have dimensions in their listings is beyond me. Listing "it fits a john deere xxyyzz model" doesn't help when you don't have that model, but the fitting might not be uncommon.
15:03 < jrayhawk> ~/win 7
15:04 < jrayhawk> whoops
15:16 -!- Urchin[emacs] [~user@user/urchin] has quit [Ping timeout: 240 seconds]
16:36 < fenn> docl: "Mesa" comes from the Greek word "μέσος" (mésos), meaning "middle" or "intermediate." In the context of "mesa-optimizer," it refers to an optimizer that emerges as a byproduct of another optimizer's optimization process. it's optimizing for an intermediate metric
16:38 < fenn> docl: just use GPT-4, it's not worth your time to use GPT-3.5 to learn incorrect things. if you have a bunch of low-skill busywork like reformatting text, that's a good use for 3.5
16:39 < fenn> docl: also you should learn how to ask questions without leading the LLM
16:39 < fenn> less is more
16:40 < fenn> btw it's pronounced "mez-uh"
17:06 < fenn> jfc i had to ask GPT-4 what the first use of the term was because rationalists are just that bad at explaining where their neologisms come from
17:07 < fenn> this is the source (yudkowsky of course, who else) https://arxiv.org/pdf/1906.01820.pdf "Mesa-optimization is a conceptual dual of meta-optimization—whereas meta is
17:07 < fenn> Greek for above, mesa is Greek for below"
17:07 < fenn> (both incorrect etymologies)
17:10 < fenn> 'Possible misunderstanding: “mesa-optimizer” does not mean “subsystem” or “subagent.” In the context of deep learning, a mesa-optimizer is simply a neural network that is implementing some optimization process and not some emergent subagent inside that neural network. Mesa-optimizers are simply a particular type of algorithm that the base optimizer might find to solve its task.'
17:18 < fenn> this is the origin of the bad etymology and probably where yudkowsky originally got the terminology from https://web.archive.org/web/20180211110720/www.gwiznlp.com/wp-content/uploads/2014/08/Whats-the-opposite-of-meta.pdf
17:19 < fenn> https://www.wordhippo.com/what-is/the-meaning-of/greek-word-74cd0e2a648f7502284fe308d6ac0aca971525f8.html
17:20 < fenn> What does μέσα (mésa) mean in Greek? English Translation: inside
17:22 < fenn> so in conclusion, it's all sorts of fucked. the jargon is less helpful than just explaining what you mean, because there are so many misunderstandings around it. people can't be bothered to introduce their new jargon and properly link back to definitions
17:43 < fenn> also docl your link to "a conversation with gpt4" was actually gpt-3.5, which you can tell because of the green swirlie icon. gpt-4 has a black (or purple?) background. i wish they would give a lot more metadata than "Model: Default"
17:47 < fenn> hmm i'm confused now because that's the same answer gpt-3.5 gives but digging around in the html it says gpt-4
18:11 < muurkha> oh geez, neurolinguistic programming
18:14 < fenn> natural language processing seems more likely
18:15 < muurkha> nope, he's citing Bandler and Grinder
18:15 < muurkha> and also, for good measure, Grinder and Bandler
18:17 < muurkha> and on p. 2 he describes "NLP" as a "philosophical system"
18:17 < fenn> GPT-4 seems to flip flop between giving me the greek etymology and hallucinating a spanish etymology when it isn't told to be smart. it always (so far anyway) gives the greek answer when using my system prompt: https://fennetic.net/irc/GPT-4_etymology_with_prompt.yaml.txt https://fennetic.net/irc/GPT-4_etymology_without_prompt.txt
18:19 < muurkha> maybe instead of asking GPT-4 you could ask Liddell and Scott on Perseus
18:20 < muurkha> well, maybe you're more interested in LLMs than in the etymology
18:22 < fenn> it doesn't like unicode
18:23 < fenn> maybe you can get closer than this https://www.perseus.tufts.edu/hopper/morph?l=me/sa&la=greek
18:25 < fenn> yeah this is an atrocious dictionary lookup interface
18:29 < fenn> it's not in the dictionary in that form. here's the form chatGPT gave me (mesos) https://www.perseus.tufts.edu/hopper/text?doc=Perseus%3Atext%3A1999.04.0058%3Aalphabetic+letter%3D*m%3Aentry+group%3D15%3Aentry%3Dme%2Fsos
18:39 < fenn> i might have more disdain for pseudoscience like NLP if mainstream psychology and psychotherapy weren't also infested with pseudoscience
18:44 < fenn> i'm guessing it went like this: yudkowsky googled "what is the opposite of meta" and landed on this page which had a nice simple answer and even offered an explanation
18:44 < muurkha> you'll note that that entry gives examples of its use as mesos, meson, mesoi, meses, mesites, mesou, messon, messoi, messoisi, mesos with an omega, mesaiteros, and mesaitatos, but never mesa
18:46 < fenn> if you understand greek conjugation, by all means judge away. i don't
18:46 < muurkha> that's because "mesa" is wrong and NLP practitioners have evidently branched out from pseudoscience to pseudoclassics
18:46 < muurkha> declension, not conjugation
18:46 < muurkha> it's a noun/adjective
18:47 < fenn> word hippo has a bunch of uses in modern greek
18:47 < muurkha> I've forgotten my declensions
18:47 < muurkha> of "mesa"? with an alpha at the end?
18:47 < fenn> yes
18:48 < muurkha> paste away!
18:48 < fenn> linked above, again for your convenience https://www.wordhippo.com/what-is/the-meaning-of/greek-word-74cd0e2a648f7502284fe308d6ac0aca971525f8.html
18:49 < muurkha> hmm, maybe it does exist in modern Greek then
18:50 < muurkha> https://en.wiktionary.org/wiki/%CE%BC%CE%AD%CF%83%CE%B1 says so
18:52 < muurkha> I still think meso- would be the proper form for -optimizers
18:53 < fenn> same
18:56 < fenn> the terminology is confusing because it leads one to believe that you're talking about sub-agents
18:56 < muurkha> mediocre agents
18:57 < muurkha> mediating agents
18:57 < fenn> it's also just ... one more "rationalist" bullshit neologism that nobody else says
18:57 < fenn> so then you have to go do the legwork which inevitably involves reading pages and pages
18:57 < fenn> and you're no better off in the end because the jargon is not used precisely anyway
18:57 < muurkha> between agents
18:58 < muurkha> during agents
18:58 < muurkha> inside agents
18:58 < muurkha> jailed agents: μπαίνω μέσα (baíno mésa, “I am overdraught; I go to jail”)
19:05 < fenn> i like abram demski's critique of the ontology. instead he proposes these two kinds of optimization:
19:05 < fenn> "Selection refers to search-like systems, which look through a number of possibilities and select one."
19:05 < fenn> "Control refers to systems like thermostats, organisms, and missile guidance systems. These systems do not get a re-do for their choices. They make choices which move toward the goal at every moment, but they don't get to search"
19:15 < fenn> "a controller with a search inside would typically have some kind of model of the environment, which it uses by searching for good actions/plans/policies for achieving its goal"
19:16 < fenn> "[its] search can only do as well as its model can tell it; however, the agent is ultimately judged by the true consequences of its actions"
19:16 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has joined #hplusroadmap
19:17 < muurkha> I don't think either of those categories includes standard optimization algorithms like gradient descent, the simplex method, or applying Newton's method to the derivative
19:18 < muurkha> or applying the quadratic equation to the derivative if it happens to be quadratic, etc.
19:18 < fenn> those are solidly in the "controller" class
19:19 < muurkha> no, because they don't make irreversible choices; all their choices get a re-do whenever desired
19:19 < fenn> hmm. i guess you could tune hyperparameters with gradient descent, okay
19:20 -!- test_ [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 240 seconds]
19:20 < muurkha> they don't work in terms of moments, they exist outside time and space
19:20 < muurkha> it's the opposite extreme from a missile guidance system
19:21 < muurkha> I conclude that Abram Demski is unfamiliar with the most basic aspects of optimization, so his opinions should be dismissed with extreme prejudice
19:22 < fenn> lol please take my paraphrasing with a grain of salt
19:22 < muurkha> you weren't quoting him?
19:23 < fenn> here's what he actually wrote: https://www.lesswrong.com/posts/WmBukJkEFM72Xr397 https://www.lesswrong.com/posts/ZDZmopKquzHYPRNxq/selection-vs-control
19:24 < fenn> if you really want to read 50 pages of discourse
19:24  * fenn goes for a walk
19:25 < muurkha> your summary seems to be correct
19:25 < muurkha> and he's talking about some classic stochastic optimization algorithms like simulated annealing
19:26 < muurkha> he just doesn't know about the deterministic optimizers that preceded them
19:28 < muurkha> which are called by the same name because they do the same thing
19:28 < fenn> i find it impossible to believe he doesn't know about gradient descent
19:29 < muurkha> or, for that matter, about the modern field of convex optimization
19:29 < muurkha> maybe he categorizes it as a "search-like system"
19:32 < muurkha> but he explicitly says, "The notion of a selection process says a lot about what is actually happening inside a selection process: there is a space of options, which can be enumerated; it is trying them"
19:32 < muurkha> you clearly can't enumerate the space of options of any continuous optimization algorithm, including gradient descent
19:33 < muurkha> nothing wrong with being clueless and writing about your own exploration process
19:34 < muurkha> but it's important not to confuse a novice thinking out loud with insights that advance a field
19:36 -!- darsie [~darsie@84-113-55-200.cable.dynamic.surfer.at] has quit [Ping timeout: 268 seconds]
19:36 < fenn> are you getting hung up on whether it's continuous or discrete?
19:36 < fenn> a selector can select from a continuous space of possibilities
19:41 < fenn> 'I don't want to frame it as if there's "one true distinction" which we should be making'
19:44 < fenn> the reason anyone's talking about this is because, supposedly, optimizing for one thing might accidentally create another thing that optimizes for a different goal, and it will go foom and kill us all
19:46 < fenn> none of this discussion has changed my initial favorable impression of, "AI risk is string theory for computer programmers"
19:46 < muurkha> you could conceivably enumerate a dense set like the rationals
19:48 < muurkha> hmm, I guess you can technically enumerate the computable numbers, and no optimization algorithm can compute an uncomputable number
19:48 < muurkha> so I guess I am forced to withdraw my objection
19:50 < muurkha> there are still optimization algorithms that don't try different options
21:22 < docl> fenn: ah, I led with the spanish word for table bit based on a prior convo with gpt3.5. gpt4 corrected that to the landscape feature analogy ([1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5C[1;5Cwhich seems to fit? looks like it was not the originally intended meaning though)
21:23 < docl> ugh, keyboard issue
21:23 < docl> anyways, thanks for the correction
21:23 < fenn> yes it's powerful at rationalizing previous context
21:25 < fenn> (this is not good)
21:25 < fenn> you can pretty much get it to say anything you want if you present it as a completion task
21:25 < fenn> first and foremost it's a text completion algorithm, that has been very slightly nudged in the direction of answering questions
21:27 < docl> but yes the link I gave was to the gpt4 convo. I guess there are some issues in the UI for shared convos still. You can sort of tell the difference based on GPT4 giving longer paragraphs in its answers. I usually use 3.5 for speedrunning and 4 for when I want further insights. 4 caps at 25 questions, and is slower to respond.
21:33 < docl> my original assumption was that mesa- was intended to be intermediate vs external, so I was surprised that 3.5 told me it was based on the spanish word for table, and further surprised that gpt4 corrected it to geographic feature (because the latter actually makes sense, it's just not uh, typical to how academical neologisms would normally be coined)
21:33 < fenn> i didn't want to give my phone number to what is effectively microsoft and unambiguously tie all my private thoughts to my personally-assigned surveillance adtech device, so i've been mooching off of other peoples' API access, which fortunately doesn't have the turn limit. it also seems to cost less with my usage patterns
21:34 < docl> er, internal vs external? something like that
22:25 -!- test_ [~flooded@146.70.195.99] has joined #hplusroadmap
22:29 -!- flooded [flooded@gateway/vpn/protonvpn/flood/x-43489060] has quit [Ping timeout: 268 seconds]
22:51 -!- TMM_ [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
22:52 -!- TMM_ [hp@amanda.tmm.cx] has joined #hplusroadmap
23:00 -!- Netsplit *.net <-> *.split quits: mrdata, otoburb, drmeister, yorick, mxz, user_
23:00 -!- Netsplit over, joins: user_
23:00 -!- yorick [~yorick@pennyworth.yori.cc] has joined #hplusroadmap
23:01 -!- mrdata [~mrdata@135-23-182-55.cpe.pppoe.ca] has joined #hplusroadmap
23:01 -!- Netsplit over, joins: drmeister
23:01 -!- mrdata [~mrdata@135-23-182-55.cpe.pppoe.ca] has quit [Changing host]
23:01 -!- mrdata [~mrdata@user/mrdata] has joined #hplusroadmap
23:02 -!- yorick is now known as Guest3564
23:02 -!- mxz [~mxz@user/mxz] has joined #hplusroadmap
23:03 -!- ANIMEX [~Malvolio@idlerpg/player/Malvolio] has quit [Ping timeout: 256 seconds]
23:05 -!- otoburb [~otoburb@user/otoburb] has joined #hplusroadmap
23:33 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
23:36 < Malvolio> .t https://lexfridman.com/neil-gershenfeld/
23:36 < EmmyNoether> #380 – Neil Gershenfeld: Self-Replicating Robots and the Future of Fabrication | Lex Fridman Podcast
--- Log closed Mon May 29 00:00:14 2023