Thought provoking write up. I think with GPT-3/Co-pilot and DALL-E 2 we're seeing the advent of a new computing interface and information paradigm. We're very early.

The old/existing way is: humans produce content (writing, images, videos, music) using software and then use the internet to host and distribute the content. As the volume of information increased, aggregators of content became more important. Google helps you find anything, but primarily text and images, YouTube indexes and recommends medium length videos, etc.

A very common problem pattern is that you find multiple sources that are _nearly_ what you want, but not quite. For example you combine two stock photos in photoshop (with some editing) to produce a final image.

The future way is: large foundational models that act as the aggregators of information _and_ the editing tools. Co-pilot is the best current example we have. Once the performance of Co-pilot becomes good enough, I will never have to leave VSCode + Co-pilot, replacing the previous pattern of Google + documentation + Stackoverflow. The indexers and aggregators will be made obsolete. This is exacerbated by the fact that Co-pilot is able to do more than just aggregating the correct Stackoverflow examples. Co-pilot can do the customization and translation of the code examples into your specific use case.

These large/foundational models are nowhere near perfect yet. Once the performance is good enough to make existing workflows easier, we'll see rapid adoption in every domain. Recommendation algorithms will look entirely different in 10 years time. The 2030 version of YouTube won't be serving users the closest match video from its library, it will be generating custom edits on-the-fly tailored to the user.

totally agree, part of why we wrote this was attempting to frame DALL-E as an early step towards something much bigger, rather than overly focussing on its current application!

What's interesting here is that as a content creator, it really does feel as if this externalizes the 'latent space' that exists in my own head. 99% of my time/skill is used in the pursuit of getting the vision of something that exists only in my mind, into a state where it can be shared with others in the physical world.

Collaborating with an external latent space and teasing certain viewpoints out of it is an interesting dance, and something that I really look forward to participating in as soon as I can get access to the tools.

I'm very curious about where the capacity for iterative refinement is/will be in the future as well. "Ok, that looks great, now can you try it with a bit more green?" etc.

Unfortunately, I can still never look at these blog posts for more than a few seconds without becoming disturbed. Any time an animal (human or otherwise) is included in the image, it's terrifying--especially the hands/hooves/paws. Granted, I'm more phobic of monstrosities than the average person, but still. The technology has some distance to cross before I can call its output beautiful.

...Never? I fail to imagine somebody finding the mouse in VR[1], the sneezing fennec[2], or the programming sloth[3] viscerally disturbing.

[1]: https://i.redd.it/rfn3io4urcu81.jpg

[2]: https://i.redd.it/5t4pphlprqu81.png

[3]: https://i.imgur.com/dzyUvX2.jpg

By "never", I'm referring to the blog posts that have many images. Some images are OK. These are mostly OK. The mouse does have too many toes on one paw, but that's not too bad. The sloth also has uneven-sized claws in a way that seems plausible but unlikely. Also not too bad. So, my "any time" may be literally disproven by the existence of the above images, but these are exceptions in my experience.

I typed:

    "a painting of a bridge, giving me the satisfaction of having painted it myself"

No such thing appeared.

author of the piece here! I think these tools are best understood as a step in an artist workflow that people will use to varying degree. Many painters use photography or sampled images to audition compositions, for example. Musicians use samples to spark an idea or set a palette for them to create within. The satisfaction of creating on a blank canvas won't change, but most artists use tools to help them realize what they want to create, and these are incredibly powerful to that end.

I think that typing to create imagery is cool, but also agree that the ability to instantly create convincing images (with those images being considered the final work) is not the most exciting/rewarding part, at least when those images are isolated as singular works to consider(like in painting) and not part of a greater narrative. This perspective was part of our motivation to frame the subject differently!

Seems trite. Like why not, "a painting that I could sell for millions"

Honestly, you could come up with some painting you would be proud of. But you'd actually have to engage with the tool and make something.

The verb paint is open to interpretation.

I downvoted this, not because I don't think there's the germ of a valid point, here. It's just that a flip dismissal doesn't help us dig into this in any way.

Creating stuff has multiple goals. I'm currently getting a lot of "satisfaction" from AI image generation. Does that rebut your point? It's hard to know...