Hacker News Reader

Comment by cl42 | Hacker News Reader

So what?

I too wouldn't know how to draw anything related to "ophanim" unaided by a dictionary.

I'd also argue that its ability to come up with symbols that look oddly like letters but aren't actual letters could be a form of creativity.

Don't get me wrong, I'm not a GPT-3/DALL-E fanboy but I don't get why this is interesting. Anyone who has worked with a creative designer/artist/architect/etc. would know that there are so many "wrong answers" when it comes to creating visual imagery. Some of these aren't limitations of DALL-E but rather misinterpretations of our own expectations.

View on HN

Failure modes often tell you more about a technology or phenomenon than successes do.

The marketing and promotion will, of course, focus on successes. Failures suggest boundaries, limits, and weaknesses.

There's a similar case for humans. Studies of optical and auditory illusions, cognitive biases, and various forms of pathological behaviours, both individual and group, would be examples of this.

Or any equipment or system. The flight envelope of an aircraft is literally the boundary beyond which its behaviour is unpredictable and/or uncontrollable. For software, limits define the extent to which applications or AI algorithms produce useful results.

How much those limits can be pushed is itself an interesting question. The observations of failures of negation (a map of the world without Europe, a bowl of fruit without apples, a man not running) suggest that DALL-E has something of a "now don't think of an elephant" problem. It's also a problem that's been observed in search contexts --- I am recalling a case where Google and/or Amazon repeatedly failed when requested to find shirts without stripes, consistently presenting striped shirts.

Perhaps semantic negation is hard?

"If you please ... draw me a sheep."

http://www.supercoloring.com/sites/default/files/styles/colo...

One thing I don't like about this article is that it's showing singular examples of Dall-E's failures, when Dall-E really produces 8 possible answers per text prompt (and more if you iterate further). You can cherry pick the worst of the batch and write an article about how bad it is just as, I suppose, you could write an article about the best of each batch and how good it is. However, the fact that multiple drafts are created is similar to the way humans work when doing creative work. So I think it's more instructive to look at the successes. I'm sure if we chose any genius in human history and analyzed their life's work, we'd find tons of failure and false starts and bad attempts. But we don't let those failures detract from their later success. However, in analyzing AI, you can quickly jump to that kind of "it has to be perfect or its fake" analysis and I think it's a flawed way of looking at things.

Okay, so the author had to iterate on a couple of examples. But look at the speed. Most of these examples would take a human a day to a week of labor, for each one. The fact you can rattle off almost any random text prompt like "draw me a walrus in the style of pointillism with a fake looking mustache" and that it can accommodate such requests in one second is absolutely stunning.

>> Perhaps semantic negation is hard?

More to the point, it doesn't do semantics at all, only syntax, if you will. It has learned correlations between images and words but it has no way to map either the words or the images to their meanings. All it knows is that this word appears often in correlation with that image. So when it sees the word, it gives you the image.

What is the word "no" correlated with? Probably, absolutely anything. Most likely "no" is as probable to appear as the description of an image with apples as it is to appear as the description of an image without apples. But "apples" is likely to describe an image with apples, even if it has "no" in it. For example "An image of apples in a basket with no handle" has "no" and "apples". So- you get apples, "no" or not.

In a sense, nobody has told it that "no" means "no".

(Also, it doesn't know that "apples" means "apples". "Apples" is only the label of some sets of pixels in a space. If you scrambled the labels so that "apples" was correlated with images of elephants... "an image with no apples" would get you an image with elephants).

Not even syntax so much as vocabulary and associations.

"No Means No" would make a great title for some future paper on AI NLP making significant progress on this problem.

Such failures are interesting, in the physicist's sense.

That it struggles with composition sometimes but not other times might indicate model has learned a particular distribution factorization which is good much of the time but makes it unable to properly account for conditional dependencies in certain queries. The issue of losing track of lots of details may also be related.

Its struggles with negation is also fascinating because that has been a thorn since the early days of GOFAI. Negation is trickier and more involved than what one might naively assume.

For those taking relief from such weaknesses, it's worth remembering that this is as bad as it is going to be and probably not for long. Model size is surprisingly small when compared to recent behemoths. That it can already do all that it does is certainly cause for concern.

Natural language has some major complexities that people don't realize. One of the most surprising one is that plurality does not means more than one.

>> Its struggles with negation is also fascinating because that has been a thorn since the early days of GOFAI.

Well, ish. Negation (specifically, negation-as-failure, i.e. failure to prove) in logic programming introduces non-monotonicity, but reasoning remains sound and even completeness is guaranteed under conditions.

John McCarthy (father of AI and also LISP) was very enthusiastic about nonmonotonic reasoning because he considered it the obvious way to represent common sense which was, for him, the one thing missing from AI systems. In plain words, nonmonotonic reasoning means that you can change your mind as new information becomes available.

But, you have a point that DALL-E's (and other similar systems') inability to deal with negation is reminiscent of older issues with expressing negation. In particular, it reminds me of datalog, a subset of Prolog that does not allow negation as failure. The motivation for that is to guarantee termination. This guarantee comes at the cost of completeness, so there are programs that cannot be expressed in datalog (but can in Prolog).

There might be something similar going on with DALL-E, or maybe it's just a simpler case of not being able to derive negation from correlations, as I point out in another comment.

>So what?

DALL-E 2 is very impressive. So impressive that many people have wondered if all commercial artists will be out of a job as soon as DALL-E 2 is made publicly available. These limitations form a strong argument that DALL-E 2 is not ready to completely replace human artists just yet. This is particularly relevant if you're an artist, or if you're in the market for one.

DALL-E is really impressive, but many believe it shows we're making progress towards AGI (the explicit end goal of Open AI), or at least it's displaying some degree of "understanding" (whatever that means).

If I ask for a red box on top of a blue box on top of a yellow box, and the model gives me an image with two boxes side-by-side (albeit a very high quality one), it's hard to believe there's much "understanding" of (a) box, (b) colour, (c) relative positioning. Likewise if I ask for six penguins and it gives me four penguins, can we confidently say there's any "understanding" of what a penguin is?

Again, that's not to say it's not mind-blowing technology/progress in machine learning. It most certainly is. But I'm not ready to give up on every other angle and go all-in on large diffusion models as the path to AGI.

Does the fact that sometimes graphic artists misinterpret commission requests imply that they are not conscious? People misunderstand each other all the time, and then play an interrogative game to try to synchronize their minds. That is, they ask questions.

DALL-E right now doesn't have the facility to ask questions, but iterating on the request statement is a bit of a proxy for it. I'd argue that this is exactly what understanding does in fact look like.

That isn't to say that DALL-E is AGI or conscious or anything like that. It's just to say that the mere presence of occasional misunderstanding is not a marker of un-intelligence. It's precisely when you are able to access the richness of symbolic reasoning that misinterpretation is at greatest risk, e.g. your calculator never misinterprets you.

what question does DALL-E ask to clear up the difference between "draw four penguins" and "draw six penguins"? how many iterations does it need to get "small red box on top of larger yellow box"?

DALL-E is very good at matching words to pictures and combining the results in novel ways. It isn't intelligence and it isn't art, and never will be. It's a tool that humans can use to accomplish tasks. That has its own merit.

Not intelligent - maybe. Not art? How do you conclude that? The output looks like art to me.

> If I ask for a red box on top of a blue box

The model has to encode those words into a fixed length vector, an architectural design decision, not an inherent requirement. For example translation systems can translate such phrases perfectly, but they don't pass the representation through a single vector bottleneck.

>Likewise if I ask for six penguins and it gives me four penguins, can we confidently say there's any "understanding" of what a penguin is?

What if you ask it to draw some Nazi paraphernalia and 3 sims?

https://www.pcgamer.com/bizarre-russian-propaganda-links-sim...

https://twitter.com/b_nishanov/status/1518637055596769282

I think some limitations are interesting in terms of tempering expectations and understanding how DALL-E "thinks" or works in order to best utilize its strengths.

Of course, I would always expect limitations based on having to "understand" anything beyond word-image associations, like floor plans or solvable mazes, which have meaning well beyond the visual. Those limitations are indeed uninteresting. Might as well try "An image proving the Collatz Conjecture"

This reminds me of when I used to put a sentence or two through Google translate and then put it through Google translate again with a different language and then paste it on my Tumblr.

It's not actually anything. It's just computer-generated nonsense that kinda looks like something.

When something is new and mysterious, knowing what it is not is just as interesting as knowing what it is.

True, but the "blue cube over the red cube" limitation is caused by a deliberate choice to maximise other abilities of the model at the expense of exact word for word attribute memory. The reason is that CLIP is encoding the text to a simple fixed length vector, the more text you put in the more gets ignored.

It's a word I use every day. The term is used in the Jewish daily prayerbook, in the prayers before the "shema" among other places:

https://www.sefaria.org/Siddur_Ashkenaz%2C_Weekday%2C_Shacha...

It would have been cool if there were a few tweets with the caveat of "tho here's what DALL-E is good at". Maybe the twitter account has showcased that separately.

otherwise so what if DALL-E has limitations on some weird obscure prompts. As I am hearing about DALL-E the first time, I am still impressed with the images in this tweet thread.

Reply on news.ycombinator.com