Imagen, a text-to-image diffusion model (opens in new tab)

https://www.wired.com/2016/04/can-draw-bikes-memory-definite...

Music, I'm afraid, appears stuck in the doldrums of small one-offs doing stuff like MIDI. Nothing like the breadth & quality of Jukebox has come out since it, even though it's super-obvious that there is a big overhang there and applying diffusion & other new methods would give you something like much like DALL-E 2 / Imagen for general music.

In practice, my guess is that even though Dall-e level performance in music generation would be stunning and incredible, it would also be tiresome and predictable to consume on any extended basis. I mean- that's my reaction to Dall-e- I find the images astonishing and magical but can only look at them for limited periods of time. At these early stages in this new world the outputs of real individual brains are still more interesting.

But having tools like this to facilitate creation and inspiration by those brains- would be so so cool.

I wouldn’t be surprised if the lack of video and 3D understanding in the image dataset training fails to understand things like the fear of heights, and the concept of gravity ends up being learned in the text processing weights.

That is what I feel personally.

People tend to really underestimate just how big these models are. Of course these models aren't simply "really really big" MLPs, but the cleverness of the techniques used to build them is only useful at insanely large scale.

I do find these models impressive as examples of "here's what the limit of insane amounts of data, insane amounts of compute can achieve with some matrix multiplication". But at the same time, that's all they are.

What saddens me about the rise of deep neural networks is it is really is the end of the era of true hackers. You can't reproduce this at home. You can't afford to reproduce this one in the cloud with any reasonable amount of funding. If you want to build this stuff your best bet is to go to top tier school, make the right connections and get hired by a mega-corp.

But the real tragedy here is that the output of this is honestly only interesting it if it's the work of some hacker fiddling around in their spare time. A couple of friend hacking in their garage making images of raccoon painting is pretty cool. One of the most powerful, well funded, owners of the likely the most compute resources on the planet doing this as their crowning achievement in AI... is depressing.

What I don't understand is how they do the composition. E.g. for "A giant cobra snake on a farm. The snake is made out of corn." I think I could understand how it could reproduce the "A giant cobra snake on a farm" part. What I don't understand is how it accurately pictured "The snake is made out of corn." part, when I'm guessing it has never seen images of snakes made out of corn, and the way it combined "snake" with "made out of corn", in a way that is pretty much how I imagined it would look, is the part I'm baffled by.

If you want the model to understand what a "nurse" actually is, then it shouldn't be associated with female.

If you want the model to understand how the word "nurse" is usually used, without regard for what a "nurse" actually is, then associating it with female is fine.

The issue with a correlative model is that it can easily be self-reinforcing.

For a one-shot generative algorithm you must accept the artist’s biases.

Does a bias towards lighter skin represent reality? I was under the impression that Caucasians are a minority globally.

I read the disclaimer as "the model does NOT represent reality".

Also, getting a random sample of any demographic would be really hard, so no machine learning project is going to do that. Instead you've got a random sample of some arbitrary dataset that's not directly relevant to any particular purpose.

This is, in essence, a design or artistic problem: the Google researchers have some idea of what they want the statistical properties of their image generator to look like. What it does isn't it. So, artistically, the result doesn't meet their standards, and they're going to fix it.

There is no objective, universal, scientifically correct answer about which fictional images to generate. That doesn't all art is equally good, or that you should just ship anything without looking at quality along various axes.

I want to be clear here, bias can be introduced at many different points. There's dataset bias, model bias, and training bias. Every model is biased. Every dataset is biased.

Yes, the real world is also biased. But I want to make sure that there are ways to resolve this issue. It is terribly difficult, especially in a DL framework (even more so in a generative model), but it is possible to significantly reduce the real world bias.

Yeah, but you get that same effect on every axis, not just the one you're trying to correct. You might get male nurses, but they have green hair and six fingers, because you're sampling from the tail on all axes.

So even if we managed to create a perfect model of representation and inclusion, people could still use it to generate extremely offensive images with little effort. I think people see that as profoundly dangerous. Restricting the ability to be creative seems to be a new frontier of censorship.

If the model only generated images of female nurses, then it is not representative of the real world, because male nurses exist and they deserve to not be erased. The training data is the proximate causes here, but one wonders what process ended up distorting "most nurses are female" into "nearly all nurse photos are of female nurses" something amplified a real world imbalance into a dataset that exhibited more bias than the real world, and then training the AI bakes that bias into an algorithm (that may end up further reinforcing the bias in the real world depending on the use-cases).

For example, Google's image search results pre-tweaking had some interesting thoughts on what constitutes a professional hairstyle, and that searches for "men" and "women" should only return light-skinned people: https://www.theguardian.com/technology/2016/apr/08/does-goog...

Does that reflect reality? No.

(I suspect there are also mostly unstated but very real concerns about these being used as child pornography, revenge porn, "show my ex brutally murdered" etc. generators.)

Chitwan Saharia, William Chan, Saurabh Saxena†, Lala Li†, Jay Whang†, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho†, David Fleet†, Mohammad Norouzi

One example would be if Imagen draws a group of mostly white people when you say "draw a group of people". This doesn't reflect actual reality. Another would be if Imagen draws a group of men when you say "draw a group of doctors".

In these cases where iconographic reality differs from actual reality, hand-tuning could be used to bring it closer to the real world, not just the world as we might wish it to be!

I agree there's a problem here. But I'd state it more as "new technologies are being held to a vastly higher standard than existing ones." Imagine TV studios issuing a moratorium on any new shows that made being white (or rich) seem more normal than it was! The public might rightly expect studios to turn the dials away from the blatant biases of the past, but even if this would be beneficial the progressive and activist public is generations away from expecting a TV studio to not release shows until they're confirmed to be bias-free.

That said, Google's decision to not publish is probably less about the inequities in AI's representation of reality and more about the AI sometimes spitting out drawings that are offensive in the US, like racist caricatures.

    Translation: we need to hand-tune this to not reflect reality

Seems to me that (as with any ML stuff, right?) it's reflecting the training corpus.

Futhermore, is it this thing's job to reflect reality?

    the world as we (Caucasian/Asian male American woke 
    upper-middle class San Fransisco engineers) wish it to be

Heartfelt answer: Yes, there is some of that wishful thinking or editorializing. I don't consider it to be erasing or denying reality. This is a tool that synthesizes unreality. I don't think that such a tool should, say, refuse to synthesize an image of a female POTUS because one hasn't existed yet. This is art, not a reporting tool... and keep in mind that art not only imitates life but also influences it.

Presumably when you're significantly predictive of the preferred dogma, rather than reality. There's no small bit of irony in machines inadvertently creating cognitive dissonance of this sort; second order reality check.

I'm fairly sure this never actually played out well in history (bourgeois pseudoscience, deutsche physik etc), so expect some Chinese research bureau to forge ahead in this particular direction.

It is also available via Hugging Face transformers.

However, the paper mentions T5-XXL is 4.6B, which doesn't fit any of the checkpoints above, so I'm confused.

I mean a good example of this is the Pulse[0][1] paper. You may remember it as the white Obama. This became a huge debate and it was pretty easily shown that the largest factor was the dataset bias. This outrage did lead to fixing FFHQ but it also sparked a huge debate with LeCun (data centric bias) and Timnit (model centric bias) at the center. Though Pulse is still remembered for this bias, not for how they responded to it. I should also note that there is human bias in this case as we have a priori knowledge of what the upsampled image should look like (humans are pretty good at this when the small image is already recognizable but this is a difficult metric to mathematically calculate).

It is fairly easy to find adversarial examples, where generative models produce biased results. It is FAR harder to fix these. Since this is known by the community but not by the public (and some community members focus on finding these holes but not fixing them) it creates outrage. Probably best for them to limit their release.

[0] https://arxiv.org/abs/2003.03808

[1] https://cdn.vox-cdn.com/thumbor/MXX-mZqWLQZW8Fdx1ilcFEHR8Wk=...

Very difficult to replicate results.

Other STEM adjacent communities feel similarly but I don’t get it from actual in person engineers much.

Moreover, the model doing things like exclusively producing white people when asked to create images of people home brewing beer is "biased" but it's a bias that presumably reflects reality (or at least the internet), if not the reality we'd prefer. Bias means more than "spam and crap", in the ML community bias can also simply mean _accurately_ modeling the underlying distribution when reality falls short of the author's hopes.

For example, if you're interested in learning about what home brewing is the fact that it uses white people would be at least a little unfortunate since there is nothing inherently white and some home brewers aren't white. But if, instead, you wanted to just generate typical home brewing images doing anything but would generate conspicuously unrepresentative images.

But even ignoring the part of the biases which are debatable or of application-specific impact, saying something is unfortunate and saying people should be denied access are entirely different things.

I'll happily delete this comment if you can bring to my attention a single person who has suggested that we lose access to the internet because of spam and crap who has also argued that the release of an internet-biased ML model shouldn't be withheld.

It's often not worth it to decentralize the computation of the trained model though but it's not hard to get donated cycles and groups are working on it. Don't fret because Google isn't releasing the API/code. They released the paper and that's all you need.

Dall-E had an entire news cycle (on tech-minded publications, that is) that showcased just how amazing it was.

Millions* of people became aware that technology like Dall-E exists, before anyone could get their hands on it and abuse it. (*a guestimate, but surely a close one)

One day soon, inevitably, everyone will have access to something 10x better than Imagen and Dall-E. So at least the public is slowly getting acclimated to it before the inevitable "theater-goers running from a projected image of a train approaching the camera" moment

And I don’t think whatever iteration of PaLM was cooking at the time GPT-3 started getting press would have looked to shabby.

I think Google crushed OpenAI on both GPT and DALL-E in short order because OpenAI published twice and someone had had enough.

I did a bit of disclaimer on my original post but not enough to withstand detailed scrutiny. This is sort of the trouble with trying to talk about cutting-edge research in what amounts to a tweet: what’s the right amount of oversimplified, emphatic statement to add legitimate insight but not overstep into being just full of shit.

I obviously don’t know that publication schedules at heavy-duty learning shops are deliberate and factor-in other publications. The only one I know anything concretely about is FAIR and even that’s badly dated knowledge.

I was trying to squeeze into a few hundred characters my very strong belief that Brain and DM haven’t let themselves be scooped since ResNet, based on my even stronger belief that no one has the muscle to do it.

To the extent that my oversimplification detracted from the conversation I regret that.

But if you’re interested I’m happy to (attempt) answers to anything that was jargon: by virtue of HN my answers will be peer-reviewed in real time, and with only modest luck, a true expert might chime in.

It’s Boltzmann and Szilard that did the original “kT” stuff around underlying thermodynamics governing energy dissipation in these scenarios, and Rolf Landaeur (I think that’s how you spell it) who did the really interesting work on how to apply that thermo work to lower-bounds on energy-expenditure in a given computation.

I said Maxwell’s Demon because it’s the best known example of a deep connection between useful work and computation. But it was sloppy.

I’ve got no interest in moralizing on this, but if any of the big actors wanted to they could put a meaningful if not overwhelming subset of the corpus on S3, put the source code on GitHub, and you could on a modest budget see an epoch or 3.

I’m not holding my breath.

For example: the high-frequency trading industry is estimated to have made somewhere between 2-3 billion dollars in all of 2020, profit/earnings. That’s a good weekend at Google.

HFT shops pay well, but not much different to top performers at FAANG.

People work in HFT because without taking a pay cut they can play real ball: they want to try themselves against the best.

Heavy learning people are no different in wanting both a competitive TC but maybe even more to be where the action is.

That’s currently Blade Runner Industries Ltd, but that could change.

I guess the concern would be: If one of these recipe websites _was_ generated by an AI, the ingredients _look_ correct to an AI but are otherwise wrong - then what do you do? Baking soda swapped with baking powder. Tablespoons instead of teaspoons. Add 2tbsp of flower to the caramel macchiato. Whoops! Meant sugar.

[0] http://slimsag.com/best-apache-chef-recipe/1438731.htm

As AI advances, a lot of people will look after experiencing life outside the digital world.

Even digital communication will not be trustworthy anymore with deepfaces and everything else, so people will want to get together more often.

Edit: for the lazy ones, yeah, digital will be a sad and heartless environment...

Considering how many of the readers of said blog will be scrapers and bots, who will use the results to generate more spammy "content", I think you are right.

I can see a past where this already happened, to paraphrase Douglas Adams ;)

A bit far out there in terms of plot but the notion of authenticating based on a multitude of factors and fingerprints is not that strange. We've already started doing that. It's just that we currently still consume a lot of unsigned content from all sorts of unreliable/untrustworthy sources.

Fake news stops being a thing as soon as you stop doing that. Having people sign off on and vouch for content needs to start becoming a thing. I might see Joe Biden saying stuff in a video on Youtube. But how do I know if that's real or not?

With deep fakes already happening, that's no longer an academic question. The answer is that you can't know. Unless people sign the content. Like Joe Biden, any journalists involved, etc. You might still not know 100% it is real but you can know whether relevant people signed off on it or not and then simply ignore any unsigned content from non reputable sources. Reputations are something we can track using signatures, blockchains, and other solutions.

Interesting with Neal Stephenson that he presents a problem and a possible solution in that book.

well, we do have organic/farmed/handcrafted/etc. food. One can imagine information nutrition label - "contains 70% AI generated content, triggers 25% of the daily dopamine release target".

I think this will introduce unavoidable background noise that will be super hard to fully eliminate in future large scale data sets scraped from the web, there's always going to be more and more photorealistic pictures of "cats" "chairs" etc. in the data that are close to looking real but not quite, and we can never really go back to a world where there's only "real" pictures, or "authentic human art" on the internet.

Again, I think we should consider "The Human Alignment Problem" more in this context. The transformers in question are large, heavy and not really prone to "recursive self-improvement".

If the ML-AGI works out in a few years, who gets to enter the prompts?

What exactly is the risk?

Still amazing that we're at a point where that's the case, they're both incredible developments.

Don't like any of the results from the real web? Well how about these we created just for you.

(Consumer demand and boredom both being infinite is another thing working against it.)

I would expect AI development to follow a similar path to digital media generally, as its following the increasing difficulty and space requirements of digitally representing said media: text < basic sounds < images < advanced audio < video.

What’s more impressive to me is how far ahead text-to-speech is, but I think the explanation is straightforward (the accessibility value has motivated us to work on that for a lot longer).

For example, corporate graphics design, logos, brand photography, etc.

I really do think inference time is a red herring for the first generation of these models.

Sure, the more transformative use-cases like real-time content generation to replace movies/games, but there is a lot of value to be created prior to that point.

Lying about ethics or misattributing their actions to some misguided sense of "social" responsibility puts google in a far worse light in my eyes. I can't help but wonder how many skilled employees were driven off from accepting a position at google because of lies like these.

Actually, I think they made InstructGPT even better at erotica because it’s trained to be “helpful and friendly”, so in other words they made it a sub.

Google is not a hobby project anymore: "don't do evil" or whatever they whittered on about back in the day.

Someone tried to say there were ethics committees etc the other day...what a bad joke. Who checks the ethics committee is making ethical decisions?

I was told I "didn't know what" I was talking about, excuse from some over-important know-it-all who didn't know what ethics was, i.e. they don't know what they are talking about.

I would love it.

[1] https://github.com/lucidrains/imagen-pytorch

Should ML/AI deliver on the wildest promises, it will be like a SpaceX Starship for the mind.

It's still an unruly 7 year old at best. Results need to be verified. Prompt engineering and a sense of creativity are core competencies.