--- Log opened Mon May 20 00:00:00 2024
00:09 < fenn> kanzure> can only a higher IQ person develop a better IQ test? if not, why?
00:09 < fenn> some classes of puzzles are much more difficult to solve than to construct, for example bongard problems
00:10 < fenn> you could imagine trying to reverse a simple hash function
00:12 < fenn> IQ tests measure the ability to solve IQ test problems, sure. but the ability to solve IQ test problems is also strongly correlated with the ability to do pretty much everything else, which is why there's only one "IQ test" rather than 50 different cognitive capability tests. it's just not worth giving a longer and more complicated test to measure a bunch of things that are correlated
00:12 < fenn> as long as you don't practice the IQ test problems, you're fine
00:12 < fenn> playing video games doesn't count
00:17 < fenn> kanzure there is a setting in discord to turn off pings, it's overview -> default notification settings
00:17 < fenn> i thought there was just no way to turn it off, because otherwise why would everything be so bad
00:18 < fenn> in server settings
00:24 < hprmbridge> harmoniq.punk> IQ test are very poor and primitive benchmarks to measure intelligence. Even LLMs benchmarks today are better in many cases. Using them for benchmarking intelligence only proves we are not prepared to evolve and discover the nature of intelligence.
00:25 < fenn> you have the same people saying the LLM benchmarks aren't "real" measures of intelligence, because the humans score so badly on them relative to the machines
00:26 < fenn> have you tried answering some LLM benchmark questions? some of them are really hard
00:26 < hprmbridge> harmoniq.punk> Can you benchmark a dolphin with IQ tests? If the answer is no, throw the shit into 🔥
00:26 < fenn> maybe dolphins are just not as smart as we would want them to be
00:28 < fenn> i'm not sure if "cognitive uneveness" is a real thing, in the absence of brain damage
00:28 < fenn> larger LLMs tend to do better on all the benchmarks
00:29 < fenn> maybe there really is such a thing as general intelligence. it seems more and more likely as we see AI conquer different domains
00:29 < fenn> obviously humans have special purpose neural hardware for audio processing, face recognition, navigation, etc.
00:29 < fenn> obviously a calculator is superhuman at simple math computations
00:30 < fenn> but the vast bulk of neural matter is dedicated to general purpose functionality
00:35 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has quit [Quit: 7C9B 2024-05-20 07:35:20:503]
00:38 < fenn> ok so there are a variety of species of dolphin with different cortical neuron counts. common dolphins have ~7 billion neurons, while bottlenose dolphins have 13 billion. humans have 21 billion, orcas have 43 billion
00:40 < fenn> "long finned pilot whales" are dolphins with 37 billion cortical neurons. none are held in captivity so it's very difficult to do IQ tests on them
00:42 -!- Malvolio [~Malvolio@idlerpg/player/Malvolio] has joined #hplusroadmap
00:43 < fenn> a gruesome aside,
00:43 < fenn> "the story of Old Tom will open some eyes and raise some eyebrows. Here, a wild Orca made what could be called a contract with a whaling community, and helped these humans capture and kill whales in return for being given the tongue and lips of the prey. This was the “Law of the Tongue” and for DECADES, these whalers had it very easy as Old Tom, with other whales alongside him, herded
00:43 < fenn> migratory Baleen whales into Twofold Bay (in Australia), and waited for the carcasses to be strung out for them to take their reward—the tongue and lips—after which the carcasses would be processed in the usual way."
00:54 < fenn> it seems like there's some taboo on literally giving dolphins and AI a standard IQ test
00:55 < fenn> GPT-4V was doing raven's progressive matrices, but they didn't provide the result as an IQ score or a form that could be converted to a score
00:55 < fenn> i guess anyone could do this now
00:56 < fenn> dolphin eyesight is not great so that test format is not what you would want
00:56 < fenn> is there a non-verbal IQ test format that uses mostly audio?
00:57 < fenn> oh i am wrong, their eyesight is fine
01:05 < fenn> https://www.wildorca.org/wp-content/uploads/2022/07/WO-Research-2022-07-25-00023.jpg
01:05 < fenn> clearly some differences there
02:04 < hprmbridge> kanzure> we could be training LLMs in recorded embodied environments, not just text vectors on input (any such vectors should probably be full traces of recorded agent activity, like internal input/thinking/ouput traces during that environment)...
02:04 < hprmbridge> kanzure> text should be presented at very high bandwidth visual input, etc
02:05 < hprmbridge> kanzure> and have comprehensible input (eg a mentor in the simulation that reads and acts out any such text)
02:10 < hprmbridge> kanzure> the tokens should be agent/agency behaviors
02:17 < fenn> there are multimodal transformers running robots now, trained in simulation with skills that translate to a real physical robot
02:18 < fenn> it turns out that reading is a very efficient way to learn about the world (who would have thought)
02:20 < fenn> imagine having a robot, but it can also read text. shocking, i know
03:01 -!- darsie [~darsie@84-112-12-36.cable.dynamic.surfer.at] has joined #hplusroadmap
03:13 -!- balrog [znc@user/balrog] has quit [Ping timeout: 268 seconds]
03:29 -!- balrog [znc@user/balrog] has joined #hplusroadmap
03:53 < jrayhawk_> if ever you want to be humbled, http://miyaguchi.4sigma.org/ has some fun IQ tests
04:51 < hprmbridge> harmoniq.punk> The number of neural connections matters. You can have trillions of neurons if they don't establish connections between them.  Connectomics mapping is the relevant aspect of brain power. Humans actually have 86B not 21B but these 86B are forming hundreds of trillions of connections.
04:52 < hprmbridge> fennfoot> i was just reading numbers off this list: https://en.wikipedia.org/wiki/List_of_animals_by_number_of_neurons#Forebrain_(cerebrum_or_pallium)_only
04:54 < fenn> 86 billion appears to be the total neuron count, not cortical neurons
04:54 < fenn> frankly i don't care how well an animal can dance
04:55 < fenn> cerebellum etc
04:57 < hprmbridge> harmoniq.punk> dance requires a crazy amount of physical perception and intelligence
04:57 < fenn> i know bird pallium is more efficient than mammal, but it seems likely that whale and human tissue is similar
04:57 < kanzure> fenn is not impressed by https://www.youtube.com/watch?v=nWfyw51DQfU&t=45s
04:59 < fenn> those are some mad leaf tossing skillz
05:24 < hprmbridge> harmoniq.punk> I believe neural nets today are dumb because relay on combination instead of relaying on permutations. 10 x 10 combinations is 10 ^ 13 but 10 x 10 permutations is 10 ^ 157
05:32 < hprmbridge> harmoniq.punk> GPUs abstract parallelism so a matrix in GPU is pure abstract. But a we can build a matrix of op amps which intrinsically performs permutations. Number of connections (permutations) is everything. With a 10 x 10 analog matrix we can outperform human brain connectomics by far. 10 x 10 permutations is more than 1 quadrillion parameters in terms of LLM architecture
05:36 < hprmbridge> harmoniq.punk> Deep Neural Nets today don't give a single fuck on its multi layer perceptrons orders. That's why it is organized in layers. But layers are just a dumb way to excite and amplify combinations to optimize for GPU capabilities.
05:36 < fenn> i don't see how 10^157 is anywhere near 1 quadrillion
05:37 < fenn> how do you get 10^13
05:38 < hprmbridge> harmoniq.punk> 1 quadrillion is 10^150
05:38 < fenn> this is news to me
05:38 < fenn> what formula results in that number?
05:41 < fenn> well anyway, 10^157 is unambiguous
05:42 < hprmbridge> harmoniq.punk> from math C (n,r) = n!/r!(n - r)!  and P(n,r) = n!/(n-r)!  you can clearly see the huge difference
05:43 < hprmbridge> harmoniq.punk> 0! = 1 on permutations
05:45 < hprmbridge> harmoniq.punk> compared with 10! x 0! = 3,6288 Million
06:41 < kanzure> "how would you feel if the person asking this question couldn't assess counterfactuals?"
07:26 < geneh2> @fenn what really gets me is that transformers can be used to increase the performance of motion planning algorithms: https://sites.google.com/ucsd.edu/vq-mpt/home
07:27 < geneh2> getting a robot arm from point A to B without hitting stuff is a difficult planning problem, but transformers are now competitive with other approaches
07:46 < fenn> optimize for energy or speed. just satisficing is barely doing the job
07:47 < fenn> i think we are really already at the point where it makes sense to have a do-everything model
07:48 < fenn> not one transformer for motion planning, one for image processing, one for speech recognition, etc
07:48 < fenn> for industrial stuff that will never interact with a human, sure
07:52 < fenn> this paper has too many unknown concepts for me to read, sorry
07:53 < geneh2> it's a measurable improvement to a problem that's existed for a long time
07:57 -!- dustinm [~dustinm@static.38.6.217.95.clients.your-server.de] has quit [Ping timeout: 252 seconds]
07:58 -!- dustinm [~dustinm@static.38.6.217.95.clients.your-server.de] has joined #hplusroadmap
08:04 < fenn> geneh2: mamba beats transformer in learning chess https://nitter.poast.org/HaileyStormC/status/1791909006828802348#m
08:06 < fenn> mamba is cheaper inference for the same quality, and scales better on context
08:06 < fenn> it also seems more grounded in control theory
08:08 < fenn> state space models
09:08 -!- gl00ten [~gl00ten@2001:8a0:7ef6:8700:19b7:5269:6d30:a6bf] has joined #hplusroadmap
09:21 < hprmbridge> pinkbearer> I like to use arpin’s definition of intelligence - “General intelligence is the ability to perform work on arbitrarily distant symbols in concept space.”
09:23  * L29Ah performs work on pinkbearer
09:23 < L29Ah> pinkbearer is accelerated
09:32 < fenn> https://www.politico.eu/article/robots-coming-ukraine-testing-ground-ai-artificial-intelligence-powered-combat-war-russia/
09:33 < fenn> “My goal,” he said, “is to make a proper version of the Terminator that will protect us and help our army.” 
09:33 < fenn> “I am sure that the question of limiting the use of AI should be resolved,” he said. “After our victory.”
09:37 < hprmbridge> pinkbearer> Man, I wish I had the money to invest in drones after the ethiopian civil war
09:38 < hprmbridge> pinkbearer> It was very obvious in that conflict that drones would be central to the next wars
09:38 < hprmbridge> pinkbearer> The 2018 war not the 90s
09:41 < hprmbridge> pinkbearer> The tigray insurgency was doing very well right up until the govt bought drones from Iran and crushed them underfoot
10:02 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
10:03 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
10:53 < hprmbridge> kanzure> https://trevorklee.substack.com/p/please-dont-take-luminas-anticavity
11:51 -!- dartmouthed [~blackunsp@li761-35.members.linode.com] has quit [Quit: ZNC 1.8.2 - https://znc.in]
11:52 -!- dartmouthed [~blackunsp@li761-35.members.linode.com] has joined #hplusroadmap
11:57 < hprmbridge> harmoniq.punk> I like Marcus Hutter's Deep Mind team: " Intelligence is the ability of an agent to successfully traverse a broad range of environments"  The more you successfully traverse the more intelligent. For me makes sense. We as humans do not outperform other species in specific environments but we can traverse the widest range of environments: water, land, mountains, air, space, digital, dexterity,
11:57 < hprmbridge> harmoniq.punk> social etc.
12:00 < hprmbridge> pinkbearer> the definition I posted is broader than yours- it is decent though
12:01 < hprmbridge> pinkbearer> to your definition the obvious question is "what is the ability that allows the agent to traverse?"
12:34 < hprmbridge> harmoniq.punk> If I build a robot that only knows to solve Rubik's Cube faster than any other robot or human on this planet does it define intelligence?   I think  Markus definition define intelligence, while your version try to define accuracy of a task implementation. It is subjective important the performance, accuracy and task implementation. Usability is more important than accuracy and you can see it in
12:34 < hprmbridge> harmoniq.punk> our society today. Chinese products have achieved usability and people choose their products instead paying more for accuracy
13:28 < hprmbridge> alonzoc> The analogy to the vogons is on point. And good god who thought it was a good idea to make antibiotic resistant bacteria which produce the antibiotic they are resistant to and then let them loose in the wild. This just seems like a opportunity to use bay area probiotic users mouths into an environment designed to select for mutacin-1140 resistance
14:09 < hprmbridge> kanzure> https://twitter.com/weidai11/status/1792144888819310825 "BTW did you know that Vernor Vinge was a major inspiration behind crypto-anarchy (see True Names and Ungoverned). So he single-handedly inspired two major movements that defined my "career": crypto and Singularitarianism. (3 if you count AI x-safety separately)"
14:11 < nsh> .w traverse
14:11 < saxo> traverse — noun: 1. (climbing) A route used in mountaineering, specifically rock climbing, in which the descent occurs by a different route than the ascent, 2. (surveying) A series of points, with angles and distances measured between, traveled around a subject, usually for use as "control" i.e[...]
14:12 < nsh> '(transitive) To travel across, to go through, to pass through, particularly under difficult conditions. '
14:13 < nsh> intelligence is not the ability to traverse a broad range of environments. a comet can do that.
14:14 < nsh> intelligence is what is says it is, discernment, picking between things
14:16 < nsh> a consequence of intelligence in an evolutionary context is broad adaption, but the we are not as broadly adapted as bacteria, or, again, rocks
14:16 < nsh> but we can choose between things on a much richer basis than bacteria or rocks
14:39 < L29Ah> your definition doesn't sound useful for intelligence measurement
16:07 -!- gl00ten [~gl00ten@2001:8a0:7ef6:8700:19b7:5269:6d30:a6bf] has quit [Remote host closed the connection]
16:12 < fenn> the trevor klee article was incredibly dumb. i want my ten minutes back
16:13 < fenn> i kept hoping for a punchline or something
16:15 < fenn> .ety intelligence
16:15 < saxo> "intelligence (n.)late 14c., 'the highest faculty of the mind, capacity for comprehending general truths;' c. 1400, 'faculty of understanding, comprehension,' from Old French intelligence (12c.) and directly from Latin intelligentia, intellegentia 'understanding, knowledge, power of discerning; art, skill, ta..." - https://www.etymonline.com/search?q=intelligence
16:15 < fenn> inter "between" (see inter-) + legere "choose, pick out, read,"
16:18 < hprmbridge> kanzure> the defamation email drama blog post was even worse
16:18 < hprmbridge> kanzure> more wasteful
16:43 -!- hellleshin [~talinck@76-230-66-1.lightspeed.cntmoh.sbcglobal.net] has joined #hplusroadmap
16:46 -!- helleshin [~talinck@76-230-66-1.lightspeed.cntmoh.sbcglobal.net] has quit [Ping timeout: 252 seconds]
16:56 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has quit [Ping timeout: 264 seconds]
17:20 -!- Gooberpatrol66 [~Gooberpat@user/gooberpatrol66] has joined #hplusroadmap
17:42 -!- gl00ten [~gl00ten@2001:8a0:7ef6:8700:19b7:5269:6d30:a6bf] has joined #hplusroadmap
17:56 -!- darsie [~darsie@84-112-12-36.cable.dynamic.surfer.at] has quit [Ping timeout: 252 seconds]
18:05 -!- gl00ten2 [~gl00ten@bl11-102-149.dsl.telepac.pt] has joined #hplusroadmap
18:05 -!- gl00ten [~gl00ten@2001:8a0:7ef6:8700:19b7:5269:6d30:a6bf] has quit [Read error: Connection reset by peer]
18:06 -!- gl00ten3 [~gl00ten@2001:8a0:7ef6:8700:19b7:5269:6d30:a6bf] has joined #hplusroadmap
18:09 -!- gl00ten2 [~gl00ten@bl11-102-149.dsl.telepac.pt] has quit [Ping timeout: 252 seconds]
18:58 -!- millefy5 [~Millefeui@91-160-78-132.subs.proxad.net] has joined #hplusroadmap
18:58 -!- millefy [~Millefeui@91-160-78-132.subs.proxad.net] has quit [Ping timeout: 264 seconds]
18:58 -!- millefy5 is now known as millefy
20:06 -!- gl00ten3 [~gl00ten@2001:8a0:7ef6:8700:19b7:5269:6d30:a6bf] has quit [Ping timeout: 260 seconds]
20:18 -!- mxz__ [~mxz@user/mxz] has joined #hplusroadmap
20:19 -!- mxz_ [~mxz@user/mxz] has quit [Ping timeout: 264 seconds]
20:19 -!- mxz [~mxz@user/mxz] has quit [Ping timeout: 264 seconds]
20:19 -!- mxz__ is now known as mxz
22:25 -!- mxz_ [~mxz@user/mxz] has joined #hplusroadmap
22:54 < hprmbridge> harmoniq.punk> A comet is not an agent, a comet is the environment. This definition is very carefully crafted.  Each word matter.
22:57 < hprmbridge> harmoniq.punk> Of course context matters, the context is made from the fundamental physical laws of our universe from which emerged life and intelligence.
22:58 -!- TMM [hp@amanda.tmm.cx] has quit [Quit: https://quassel-irc.org - Chat comfortably. Anywhere.]
22:58 -!- TMM [hp@amanda.tmm.cx] has joined #hplusroadmap
--- Log closed Tue May 21 00:00:01 2024