lahwran

I use talon regularly and in fact had it active while browsing hacker news. this comment is written and submitted with talon!

like Her, but there could be many or there could be one

scienceclic also comes in an english edition. hands down the best visualization of general relativity I've seen.

https://www.youtube.com/watch?v=wrwgIjBUYVc

- https://www.youtube.com/c/SimonsInstituteTOC simons institute (UC Berkeley): advanced academic math, cs, and other interdisciplinary seminars, streamed live

- https://www.youtube.com/user/IPAMUCLA (UC LA): institute for pure and applied mathematics: more advanced academic math and cs seminars

- https://www.youtube.com/user/NaSESYNC national socio-environmental synthesis center: interdisciplinary science seminars

- https://www.youtube.com/channel/UCcr5vuAH5TPlYox-QLj4ySw alan turing institute: variety of advanced cs

- https://www.youtube.com/watch?v=ENpdhwYoF5g Schwartz Reisman Institute: inter-agent friendliness/game theory

- https://www.youtube.com/channel/UCCcrR0XBH0aWbdffktUBEdw mutual information: visual explanation of ML fundamentals

- https://www.youtube.com/user/JimBobJenkins game theory and international relations

- https://www.youtube.com/channel/UCLB7AzTwc6VFZrBsO2ucBMg robert miles: "ai safety"-crowd ai safety videos. don't let this be your only perspective on ai safety, the ai safety people are great but their whole community has a shared anxiety disorder. heed their warnings but don't have the meltdowns they accidentally encourage; just imagine yourself petting the ai safety researchers on the head and going "there there, the ais will be friendly because of your work, thanks for getting me up to speed" and listen thoughtfully. emotional tone warning aside, I do like this channel. just not as much as the simons institute's videos on ai bias, alignment, safety, objectives, etc etc etc.

- https://www.youtube.com/user/PaulHBeckwith large scale climate science dude who I think isn't crazy but I'm not sure

- https://www.youtube.com/channel/UCFXBh2WNhGDXFNafOrOwZEQ/vid... oxford vgg continues to be an impressive vision group

- https://www.youtube.com/c/NormalizedNerd/featured same kind of stuff as mutual information

- https://www.youtube.com/playlist?list=PLbg3ZX2pWlgKV8K6bFJr5... art of the problem makes very good visual explanations of middling to advanced cs topics

- https://www.youtube.com/user/TheRoyalInstitution popsci talks that typically get fairly advanced. better ted talks, in a sense.

- https://www.youtube.com/channel/UC8aRaZ6_0weiS50pvCmo0pw institute for advanced study is not messing around with their name

- https://www.youtube.com/channel/UCwG9512Wm7jSS6Iqshz4Dpg ACM SIGPLAN: formal methods, formal verification, formal languages, etc

- https://www.youtube.com/c/Cirm-mathFr another advanced math channel I watch too little to remember details about. browse it yourself if you want

- https://www.youtube.com/c/RationalAnimations animations about the future and stuff

- https://www.youtube.com/c/Sevish exceptionally weird music

- https://www.youtube.com/c/JordanHarrod ai and ml stuff

- https://www.youtube.com/channel/UCX3XfA9qjWjymue2I_hcW1A formal verification and stuff

- https://www.youtube.com/channel/UCUBpU4mSYdIn-QzhORFHcHQ more formal verification stuff

- https://www.youtube.com/channel/UC9infsKo33_2LUoiqXGgQWg society analysis and stuff

- https://www.youtube.com/c/DanWorrall really good audio dude

- https://www.youtube.com/c/12voltvids small electronics channel I don't see mentioned here yet

- https://www.youtube.com/c/THUNKShow STEM video essays and explainers

- https://www.youtube.com/channel/UCrCTC5_t-HaVJ025DbYITiw alice cappelle: video essays I guess idk

- https://www.youtube.com/channel/UCecF2icZlEIJ__9XS6woPGw zoe bee: how to talk to angry strangers on the internet and stuff like that

- https://www.youtube.com/channel/UC4V_jMdRbbTrmBVJB6FDzgw unlearning economics: a dissenting perspective on economics; don't skip the other economics channels just because you watch this one, but it combines well with the others imo

- https://www.youtube.com/channel/UC_7FDEMkBBWWcol7QPzToaw one dude's commentary on native history; don't skip other commentaries on native history just cuz you watch this one, but again, combines well imo

there, some channels. there are a ton of channels already mentioned here; it's great having an index of them, but remember to browse them yourself and skim some videos.

have you seen this critique video of school of life? https://www.youtube.com/watch?v=JlkJJygIoVU

it means a test that was supposed to check if hbo max's software can send emails correctly was run with their real email authentication somehow. I'm so curious how they managed to send it to so many people!

you're correct that this sort of persistent world modeling is needed for self driving cars, but from what I've heard from friends who work in the industry, both cruise and waymo have it. they're very far from using a plain CNN on their video cameras, they've got depth mapping and such and carefully constructed software making use of the perception data to model how the world will change and react to that. idk if it works well, but they definitely know they need it and are trying.

that said, I've driven a tesla on autopilot, and holy crap is was so incredibly bad. I'm optimistic about self driving cars in general, but not about tesla's. it will frequently lose track of the road lines at night and fail to make turns, suddenly beeping at you that you're in control now, with no warning! I only ever used it like cruise control, but I can't understand how anyone driving a tesla would dare use the tricks that allow bypassing the restrictions that prevent taking your hands off the wheel.

hmm, that explains the color, but it doesn't explain the coherence. it is cool though, aerogel makes a lot of sense to use for this.

huh that's really interesting about fire's emission spectrum!

regarding 6400k - even assuming the sun's spectrum matched the black body spectrum of its surface temperature, its surface temperature is closer to 5770k. but even taking that into account, its spectrum doesn't quite match 5770k in space - and the atmosphere changes it even further. https://en.wikipedia.org/wiki/File:Solar_spectrum_en.svg

I really don't know how much of that change us humans can perceive, but my thinking is that 5000k is probably closer to the center, at least. it might be more of a saturated color, though, since the spectrum is more pointed vs the very flat but spikey spectrum in that plot.

psa: you can get this thing, which isn't the most precise ever, but lets you see this spectral information about lights for cheap: https://www.amazon.com/EISCO-Premium-Quantitative-Spectrosco... - I got one and WHEW CFLs' band lines are really obvious. Also, I feel kind of tickled that I saw the band gap on LED lights before seeing a description of what it is in the source for the spectrum plots in this article (the source being https://commons.wikimedia.org/wiki/File:White_LED.png ).

I'm a bit miffed it seems so hard to find lights that don't have this problem, though. maybe we can improve it by getting the word out that these cheap little diffraction devices can give you a pretty good approximate reading of the smoothness of the spectrum of a light source. Hmm, I just realized I have yet to take this to home depot...

I'm really curious about those MIT incandescent bulbs. If they worked well and haven't been brought to market, it's possible that contacting the people involved in creating them could have good results in making them happen. Perhaps they could be convinced to prioritize it if a case can be made that it can have a significant positive impact on the world?

What are your thoughts on how to prevent the entire class of vulnerability from being able to happen again?

So they invited the public to come watch in a venue in the north bay a year or two ago, an hour ish north of sf, and I went - it was, eh, okay. They wanted the crowd to perform cheering and stuff, it was really more like being an extra. The fights were cool, but it was pretty optimized for being a tv show and not very optimized for being a tv replay of a live show. Shrug, seems fine for what it is, just for what it's worth.

Are you imagining this running inside a cuda kernel? If yes, the problem with doing anything of the kind is that, even if you run your kernel with one warp and mask all but one thread, you still then need round trips to the cpu to do io, and you need to dynamically load code, both of which the gpu is quite bad about.

If no, it is probably doable if a bit hard to generate kernels on the fly.

No, there are more dimensions than just data efficiency; what I would say is that humans are just about as data efficient as you could possibly hope for _at that wattage_. It's easy to do better than evolution at something it wasn't trying to do, I agree - and this is the point where, as I thought about it, I realized I do actually think there's some concern. But I don't think we'll be able to do it on one gpu, because:

- human evolution has spent quite a while in an adverserial environment - the smarter you are, the more you win - a recent finding of the neural network research is that local minima are kind of not a problem in very highly dimensional spaces, as long as you have a problem that is smooth and has optima. If it has any optima, then in very high dimensions, there's probably always something you can change that will keep you moving towards the optimum. while evolution may have gotten stuck in a general class of architectures - neural ones - it seems very much like there are many dimensions along which it can change the brain, and that changing the genes for it slightly will change its performance slightly (fsvo slight). - evolution, in species that have learning systems in the first place, optimizes for intelligence per watt. energy is very costly in the wild, and so finding algorithms that work well with low power is very important. It so happens that algorithms that minimize power usage are theoretically tied to algorithms that compress well, but the key thing is that evolution has had a crapload of optimization time for tuning the brains of mammals in general, and then humans got in this runaway optimization process - which seems to have made us smarter primarily by making our brains use more power for the relevant parts.

I definitely think you could do better, I just don't think you're going to do it with a paradigm that looks vaguely like the brain, because if it looks vaguely like the brain, evolution probably passed it up on the way to the general architecture that mammals use, and the specific one humans use. Possible exception for gradient descent and weight sharing, because those would be difficult to implement in the brain, but that doesn't give you results hundreds of times better, and it's not even clear the brain doesn't do that - hinton has made the argument that it could.

the key thing here: if we make an agi with neural networks (which at this point is almost a for-sure thing), then going beyond human level on one gpu will be a very difficult research task, and take it a lot of learning to figure out how to do. Which means we'll get a chance to control it using less formal mechanisms than miri demands out of their work.

(I don't think miri's stuff will be done in time to be useful to anyone.)

oh my god yes. I'm frequently creeped out by how much my various employers have wanted me to be TOTALLY AND ENTIRELY on board with their mission. like, I like making awesome software, and I like customers enjoying it, but plz no I do not want to devote my life to x thing just because it's both fun and gives me money. If I'm going to devote my life to anything, it's going to be doing something like building computational models of the genome or something.

primarily in person, yes. presumably primarily from people who frequent less wrong. I don't, so I don't really know.

So as I was writing a reply, I found that my opinion doesn't actually differ in "do we need to do a bunch of stuff"; it's more that I don't think miri is taking a useful approach. After talking to a friend who is also in ml about this really, I think the key point I should make is that it's more a matter of engineering control tools that ensure we can spy on its thoughts. If we can build an agi, we can also build safeguards that it doesn't realize are there until too late.

Anyway, here is the original comment I was going to write. You can read it and extract your thoughts; I'm fairly confident I got the theory right, but I've reduced my confidence in the point/counterpoint. Regardless, I definitely think miri's position is unreasonably extreme, and this is a pretty ok explanation of why.

https://gist.github.com/lahwran/99d84c3f8461ece9153f

Eh... Sort of? The two organisations together would form a cult, but the community split I mention makes it a bit confusing. Overall, I agree that miri's level-of-cult is too damn high.

As someone who went, I agree. I think most people who do it right now consider the high price to be a donation to help them scale, rather than an actual product-for-money trade. I'd pay $500, maybe, if I thought the marginal benefit they'd get from my going was negligible. That's how much conferences of that length usually cost, anyway. Also, it's totally right that the techniques aren't exactly the point - you don't go to a tech conference because you can't read about the things presented there elsewhere, you go because you won't focus enough on them. But it's definitely not 4k of value from the workshop, it's 3.5k of giving them momentum to refine and scale the org, and .5k of actual experience.

Ugh, there's a lot of argument about that in the cfar alumni community. Some folks take it for granted (why??? Take things for granted about such an uncertain subject???) that we're just doomed unless you Give Miri Money(tm). Those folks tend to be pretty good about actually carrying out what is reasonable behaviour in most other ways if only they were right about that one thing - if it really were such a big deal, you'd want to not ignore it.

Meanwhile, another part of the alumni community actually understands the theory behind ai and machine learning, and those folks end up in arguments frequently with the first category about the topic.

The reason you hear about it is the first category is a pretty panicked and hopeless group - for the people who actually believe yudkowsky's "recursive algorithmic improvement" to be able to give large improvements, they generally think that humanity "loses by default" if they do nothing. So they tend to be very, very into recruiting. Thankfully they're not so nuts about it that they'll never change their minds, the problem is it takes a lot of explaining to get the theoretical basis for why the recursive self improvement thing isn't actually as scary as they think it is. No, it's not going to take an hour as soon as an ai is built, learning is hard, and humans are freakishly good at it.

Computers will beat us at data efficiency eventually but it's gonna take a while, and current machine learning is better at being data inefficient but getting good results from the large amounts of data. And the best you can do isn't good enough to make miri's monster - unbiased, maximally data-efficient Bayesian inference doesn't actually fit in the universe in either a time or memory sense if you try to build a full ai out of just that one thing. And approximating it is, you guessed it, less data efficient.