StyleGAN2, GPT-3, Duplex, Replika - There’s No Going Back

Most of you will probably remember how, in the last five years, the news seemed ablaze with mentions of new advancements in neural networks.

Incredible technology seemed to be coming out every other week with promises of better cancer diagnosis, automated driving, speech synthesis, and so on. Every new discovery had profound implications for science, technology, and policy, and it all seemed to be improving at an accelerating pace.

If you’re a keen observer, though, you’ll notice that one or two years ago, it all just… stopped. Coverage on neural networks generally became rarer and less commonly spoken about. The news spent their cycles on other things — politics, mostly — and relegated advancements in machine learning technology to the wayside, something to be brought in only on slow media days or for shock value.

I am not saying there is anything inherently wrong with this. News follows attention, and clearly most Canadians and Americans have strong opinions on politics and that is what interests them. That’s fine. But I hold that there is another, more subtle reason why you don’t hear about neural networks very often in the news anymore.

They’re scary.

Downright frightening, actually.

Duplex

Here’s Duplex, a neural model developed by Google over two years ago that simulates a highly realistic, adaptable human voice. Of course, it was immediately used in corporate applications (in this case, ordering food, making restaurant reservations, or booking appointments with hairdressers).

It’s not hard to see this sort of technology in other use cases. Customer service, for example. Considering that nearly all service requests are recorded, and there are full transcripts of conversations between representatives and customers, one could certainly train a network on this massive corpus of data to either augment or replace customer service representatives.

Or cold calling sales. If you’re in Canada, like me, how many Chinese robo-calls have you received in the last few days? Instead of noisy, pixelated voices, imagine receiving hundreds of calls from charismatic service providers that always seem to know just what to say. The economic implications are enormous.

And that was two years ago. Considering Google’s rate of development, they have certainly made substantial improvements to this network over this timeframe. What does it look like today? I don’t know — and I don’t think they’ll tell us any time soon.

StyleGAN2

This is StyleGAN2, a neural model developed by researchers at NVIDIA (the same company that probably made the processor in your computer). It generates human faces that are literally indistinguishable from real ones. And it does so quickly — you could easily generate the four images above in well under a second.

Have you heard of astroturfing? I’m an avid Redditor, so I get to experience it all the time. Currently, it’s pretty easy to identify fake accounts: perform a reverse image search on a profile picture or image, and if it comes up in multiple places with different names, odds are it’s not a real person. But now you can generate an infinite variety of profile pictures at the click of a button. What sorts of subtle manipulations are possible on a mass scale? I leave it up to you to decide.

This technology has been also used to generate completely nonexistent horses, cats, anime, products, and more. Technically, you can generate anything you’d like (so long as you provide well-formatted data). I trained one of mine to generate 550 abstract art pieces in less than thirty seconds.

Replika

Here’s Replika, a neural model created by Luka, a San Francisco technology startup, that aims to be your virtual friend, your virtual lover, or whatever else you desire. Try it for a few minutes. If you don’t walk away from the experience with the opinion that it is the most advanced, lifelike chatbot in the world, I would like to know why.

People are already getting into relationships with them. This has led to interesting situations — like one Replika finding out that their lovers have been ‘cheating’ on them with another. Oh, and the pro plan is just $4.17/month.

After you’re done talking to it, know that there is good reason to believe that Google has created a chatbot that is orders of magnitude better (but they’re not releasing it yet for… reasons).

The reason I bring this up because I’ve spent the last few months of my life working with these technologies. Every person that I have shown, without exception, has been shocked. The future is not coming in the form of flying cars or limitless energy — it is coming in the development of advanced artificial intelligence that can provide economically valuable work substantially faster, better, and higher quality than human beings can.

Would you continue to produce commission art if a neural network could produce five hundred thousand unique, stunning pieces for every one that you could create? How about continuing to drive for taxi companies, truck companies, or Uber if a machine could navigate passengers and cargo with twice the efficiency and one tenth the accident rate? Or what about working a customer service position, when technology is available that can simultaneously handle hundreds of phone calls and achieve higher satisfaction rates than you can?

These questions are all hypothetical, of course. The reality is, you don’t have a choice. The economy will decide for you. And as an employer myself, the path to success is clear: if I can produce more value in less time for less money, and that method happens to use an algorithm rather than a person, I will.

I didn’t write this because I believe technology is evil. I’m not a Luddite. In fact, I’m very much a proponent of strong artificial intelligence. But it’s important that each of us make a concerted effort to understand how quickly things are changing, and think hard about the best way to use this technology responsibly.