l33tman

You can use the dyson structure for other things.. there is a (sci-fi'esque) suggestion called a Matrioska Brain I think, where the entire dyson structure is made of computational substrates - it could be conveniently powered by the solar radiation but also by fusion. Given the track of AI and realistic world sims, everything might end up "living" inside one of those structures..

You would have to protect it pretty decently though..

Hm actually I remember I read in the studio's reply to this, that there weren't 750 active artist names at the same time, this was over many years and they release under some names, then let it go and do 50 new names etc. Still.. it's some kind of diversion of the consumer's expectations. It's one thing to see a pseudonym, but in this case, it was under "normal sounding" names.

Also, the composers' names were faked as well, so there were 750 artist names and 50 composer names..

In this case it's not pop music, it's instrumental ambient music and childrens tunes.. Very particular set of playlists, Spotify-owned, that if you get into them you're essentially guaranteed millions of streams (as those playlists have tens of millions of followers that I'm sure just have them on repeat often - guilty here as well..)

Aside from the sketchy issue with faking 750 artist names, apparently the music is good and has good artistic intent within the genre.

According to the article, you can do deals where you trade away part of your royalty for a higher placement on those Spotify-owned playlists, which in his case seems to be the right business choice to make by far.

You can apparently trade off 3/4 of your royaltys for a higher placement on the coveted Spotify-owned playlists of different types of ambient music, childrens songs etc. According to the article. But stuff like that is not very transparent I guess..

How would you otherwise define value, except the value that someone is willing to assign to a share of a company or goods when purchasing?

Value can go up and down though..

I think the MoE models are trained together just like any other network though, including the dispatcher layer that has to learn which "expert" route each token to. Perhaps you could do some kind of technically worse model architecture that is trained separately and then a more complex dispatcher that then learns to utilize the individually trained experts as best as it can?

I would buy THAT for a dollar!

Competing Interest Statement N.R.B., J.L.W., R.J.R., A.J.B., C.W., P.J.Y.L., B.H., and D.B. are co-inventors on U.S. provisional patent number 63/607,651 which covers the computational antibody design pipeline described here.

Greed is good. Especially in fundamental medical research it seems.

Yeah it's a spectacular transplant procedure, though it's really not that modern medicine, it has been developed since the 50s! The step where you irradiate or apply chemo to kill off the existing bone marrow had to be done at first because the patients had leukemia so you had to do this anyway. An interesting thing is that if there are still cells left in you, they will be wiped clean by the new transplanted immune system in a so called graft-vs-host response that also sounds like a horror-movie concept :)

Cerebras: -Hold my beer

Weird, I've had Samsungs flagship phones since 10 years back and never had any app on them that would be considered bloatware. There are a bunch of low-key note-taking apps etc that you can just ignore. Like another poster wrote, maybe it's a US-thing?

I actually use the Samsung web browser, because it allows ad-blocking out of the box, which Chrome on Android won't allow...

I've always used gboard as a keyboard (googles keyboard) and disabled the options to get really predictive and smart as I'm sure that makes it learn more closely and potentially send back data to google etc.

It apparently wasn't mandatory - you could ignore patching it if you kept doing the "reboot at least once every 22 days" directive from the AD 2016...

Avoiding to get into a "which car brand is best", I'll just say that I think there is simply a difference in the "DNA" of the car brands for how they approach the driving experience. It takes many many years and iterations and care about driving a car to produce a good car UX and some brands you can just immediately conclude that the CEO or the product managers don't care about driving and probably have never driven the specific model themselves, while in some brands you really feel they are a "driver's car".

This goes for tech in general and not just cars btw :)

It seems a bit weird to quietly hope that the money you put in an organization with the explicit goal of being a non-profit, would give you direct monetary returns though.. Maybe they hoped for returns in other ways, like getting some back-channel AGI love when it finally became conscious? :)

The Narrative Clip was created partially for this (giving people a photographic memory of sorts), and it was used by a lot of people with alzheimers as it turned out you can remember stuff better if you review every day's memories in the evening, or something like that.

That use-case seems inefficient to solve like that in the long run as well, like if you really would have to use a million tokens to do every small query you require on your data it would be prohibitively costly except doing as an experiment.

Thx

Yes exactly its kind of implied, and not trying to be rude.. it would help if the person posting the paywalled link also posts an archive link of course!

Yeah but we're not talking about LLMs here but vision transformers, which don't use the same type of token vocabulary to produce embeddings from the input as the LLMs do. The pixel data is much more dense than a few characters is, per token.

I looked it up - the original ViT models directly projected for example 16x16 pixel patches into 768-dimensional "tokens". So a 224x224 image ended up as 14*14=196 "tokens" each of which is a 768-dimensional vector. The positional encoding is just added to this vector.

This blog-post has the specific numbers, which makes it a bit less abstract than in the original paper: https://amaarora.github.io/posts/2021-01-18-ViT.html

But they are not a "single integer" either as in, like a byte... I don't have any good examples but I'm pretty sure the tokens are in the range of thousands of dimensions. It has to encode the properties of the patch of the image it derives from, and even a small 40x40 RGB pixel patch has plenty of information you have to retain.

Paywall

I'm not a layperson in this subject and I get confused. :)

Ah my mistake, unfortunately I can't correct my comment now

Getting OT, but what is the deal with all the completely different moderation guidelines (that amass to like 20-30 weird rules and exceptions) for every subreddit. I find it almost impossible to participate (except just adding a comment here and there).

For example, I wanted to post a funny Risitas youtube vid I made (you know the Spanish comedian with that laugh...), and couldn't find a single usable "funny" subreddit. Some banned youtube content completely. Some banned "video memes", some banned X and some banned Y... all of them had slightly different guidelines and you immediately got an insta-splurge from a bot-mod if you tried posting. Some required you to prefix every post subject with some code word. I had to give up eventually and post it on some super-small subreddit instead that accepted anything.

Maybe we'll come full circle back to the 80's and 90's movie scene where you had movie stars and directors who people would go to the movies to see, and for each of these you have about 100,000 2 hour AI generated movies with bland content that will have 100 views each. If that's the outcome, it's not as dystopic I think...

I mean already now you have an incredibly large fraction of sh*t on spotify, netflix etc. Will it matter if it goes up another factor of 1000x?

Meanwhile hopefully the really creative and interesting content creators will leverage the AI stuff to further take the art forward. They have much more ways to test out alternatives and scenes before committing to a $250M production shoot for example..

sama sat on twitter during that day and rendered prompts from people on the fly, so those weren't heavily cherry-picked. I didn't watch most of them and they had some bugs for sure, but if they had had to cherry-pick like 1 of 100 for the blog demos he would never have sat and done live renders like that.

Yeah text-prompting only won't replace production work, but it can surely accelerate and boost pre-prod work, maybe doing prelim shots for a storyboard, creative ideas for the costume dept... lots of cool things.

I think for prod work it needs what the auto1111 plugins have given SD - ControlNet etc, where you have a lot more control over the generation and the text prompt is just a small part of it.

I can't know if you've actually used these tools, but it requires a pretty high level of creative mind to get them to produce the content you're looking for. Maybe you as a user of an LLM you don't need to be creative in the writing of words for example, but you instead need to be creative in how you control the tools and pick the right outputs, feed it back, copy/paste/cut it, change stuff, extend it.. and the same with the image generators. There's a HUGE amount of creative accessories around them to manipulate and steer the process. There might be less creativity needed with the pen, but it's needed in other ways.

That does not follow. What inaccuracies are you thinking about?

It reflects the fact that the measurement apparatus, environment in general and the observer in particular are all also part of the "wave function". Until these are linked to the system to be observed (like the proton), the system has to be analyzed as the full set of possible configurations of its constituents.

After you link it up with the apparatus, it is pulled into the system as a new part of it, and so on. The more stuff you pull into the system, the less number of different configurations you sum over, eventually you end up with a single configuration with an observer that has a clear measurement result in her mind. This process is what is called "wavefunction collapse" in old physics texts (and a lot of modern pop-sci accounts).

What sets this apart from other purely probabilistic theories is that it's non-local, the entire linked system from the proton to the measurement apparatus to the observer has to be taken into account in the calculations if you are to be thorough. In local theories you can separate the parts of the system and handle them separately, like you could say "there is a 1% probability that a charm quark pops out in the upper left part of the proton every second" etc. You can't really do this in QM/QFT, and this is what in essence leads to all of the counter-intuitive results and confusion..