In the YouTube companion piece to this blog post (check it out if you haven’t seen it yet) I go into detail about how easy it is to use Open AI’s Dall-E 2 software. It’s as easy as uploading a picture or entering a text description of whatever zany thing you want! This software goes far beyond having a bit of fun however: it is the leading indicator (the canary in the coal mine if you will) of massive disruptions to come.
This software goes far beyond having a bit of fun
OpenAI, co-founded by Elon Musk (of Tesla, SpaceX, Neuralink and Boring Company fame) has for the past four years employed some of the most talented AI engineers on the planet doing seemingly miraculous things. From GPT-2, to GPT-3, to Dall-E, to Dall-E 2, and soon to GPT-4 and other projects, OpenAI has pushed the boundaries of large language models—better termed massive parameter models now since they clearly work on images as well as language. But this has mostly been in a research-type situation where results are “cool” but won’t be replacing jobs or making OpenAI money any time soon.
OpenAI has pushed the boundaries of large language models
Here’s the thing: that “any time soon” is suddenly upon us. Dall-E 2 is now commercially available and it’s astoundingly good! As the images you see alongside this blog show, Dall-E 2 can produce some remarkably wonderful, witty, and Creative output (see my previous blog about large models and creativity). What’s happening behind the scenes however is even more critical to the long-term prospects of products like Dall-E 2 and companies like OpenAI, as well as their competitors, like Deep Mind, and you as an employer and an employee.
Here’s the thing: that “any time soon” is suddenly upon us
The three biggest challenges to the progress of AI (or machine learning) are: better algorithms; better hardware that works faster and more efficiently; and more and better data. The better algorithms part has been handled, it seems, by transformers. These almost magical neural networks can continue scaling and getting better long past when other deep neural network architectures give up. Hardware is doing what hardware does: it’s getting better and better as fast as people (and AI) can crank out new chips. But what about Data? How are we getting better, and better quality data, on which to train these giant beasts?
These almost magical neural networks [transformers] can continue scaling and getting better
That is where we see the true brilliance of Dall-E 2’s rollout. Let humans—many millions of humans—request of Dall-E whatever their hearts desire (excepting things like famous people, banned speech, or pornographic images, thank goodness), and Dall-E 2 creates it. THEN we humans provide critical feedback by telling the program which images we prefer. We can save the images or even download them, and that “thumbs up” gives OpenAI crucial data about which results are the best. And guess what? We pay for this privilege.
But What About Data?
Think about what Tesla has pioneered: having people pay to get Tesla’s Full Self Driving (FSD) beta, and then use it which provides incredibly valuable data that helps Tesla make their FSD product even better. The same thing is starting to happen with a purely AI driven product in Dall-E 2. As Ash Jafari has noted (see here and here), this is the dawn of the AI job suite era, and if the past is any indication it will ramp up very very quickly!
This is the dawn of the AI job suite era, and if the past is any indication it will ramp up very very quickly
What this means for you as a consumer is that you are going to have lots more choices of very creative work, like graphic design, writing, marketing, etc., all much less expensively and much more quickly than before. What this means for you as a worker is that AI might well be coming for Your job soon: whether you do basic content writing or are the CEO of a large company (ahem, Elon Musk), you can and might well be replaced by an AI bot that can do the job better and cheaper. And what this means for society at large is that the compression of money into fewer and fewer companies will continue.
With a data flywheel like OpenAI is setting up right now, they will have an unassailable lead in text-to-image content creation. They might also capture concept to content (marketing, or fiction, or blog writing—hmm!), or other areas too. Or one of a few other companies might grab the prize instead. Whoever captures each area, the winner will take most of the spoils, which means a few companies will making the lion’s share of the profits.
All Images except for the AILon Musk thumbnail are credit to OpenAI's Dall-E 2, used under license. AIlon Musk Thumbnail is credit John Gibbs
Do you see the disruption that Dall-E 2 is showing us? It might not seem like a big deal today, but tomorrow you’ll notice! The best you can do is be aware of what’s happening, invest wisely (do your own research of course), and above all keep up with what’s going on in the AI world. It’s a brave new world, and terrifying or not, it certainly won’t be boring!