The AI is coming

May you live in interesting times, goes the ancient Chinese proverb. Recent developments in artificial-intelligence, especially image-generation, are putting that curse to the test.

In the last year we’ve seen some incredible breakthroughs in image-generation with the release of new algorithms: Dalle 2, Midjourney, and Stable Diffusion.

Stable Diffusion is especially interesting because it's free and open-source being funded by donations. It’s crazy to think it was released at the end of August, less than a month ago, and it already has more than a million signups. The speed of adoption is breathtaking.

Midjourney has their own algorithm. They seem to use a combination of Stable Diffusion and some other black-magic sorcery behind the scenes. A while back I started hanging out on their Discord forum, out of pure curiosity, watching people experiment with their new magic powers. At first the images were a little crude, like a ten-year-old's drawings. But then Midjourney released their new beta and all hell broke loose.

Suddenly these images came to life in beautiful vibrant detail. Stunning vistas. Alien worlds. Watercolors, oils, Monets, cartoons, cyberpunk - every style under the sun. And portraits, something that other algorithms have struggled with, Midjourney now generated with ease.

(As a weekend project I created a little Instagram bot that posts some of the community favorites.)

And then Midjourney turned the beta off because… how do I put this… things got a little porny. The algorithm itself seemed to have some disposition towards nudity, not that people needed any encouragement. (They have since “fixed” this issue and released an update).

But the cat's out of the bag. It is now possible to generate a stunning lifelike image from scratch in thirty seconds by typing a prompt. Truly incredible.

Another thing I had not predicted was independent teams, like Stable Diffusion, entering the race with open-source algorithms. Training one of these algorithms costs a lot of money. You'd think that would be out of reach for anyone but the richest tech companies and governments. But somehow here we are - we now have free models available to everyone.

Adding open-source to the mix is quite a bombshell. It really changes the incentives for companies like OpenAI and turns this into an all-out arms race.

The backlash

Thirty years ago the future we imagined was robots doing all our household chores, making the bed, washing up - you know, the boring stuff. This was something we could all get behind automating. What we didn't anticipate was that our most sacred pursuits, writing, art, and music, were to be first on the chopping block.

Recently, we engineers got a taste of our own medicine with the release of GitHub's Copilot, an algorithm that auto-completes our programs. Simply type the start of your code (or even just a comment) and Copilot suggests the rest.

I've been using Copilot for a few months now and I'm now finding it difficult to code without it (like when on a plane), because the suggestions are that good. I'd say maybe 40% of the time Copilot suggests exactly what I'm intending to do - and there have been a few instances where its results were so uncannily good a shiver ran down my spine. I can only imagine what it's going to be like a few years from now.

Up to this point, AI has been humming along in the background, quietly improving our lives. It hasn't replaced our jobs - yet - just made us more productive. It has augmented, not retired us. But is that about to change?

There has been quite some backlash from the engineering and art world over these AIs.

The stages of grief, whenever a new AI technology is introduced, seems to be:

1. "Pssht, that's not any good. My job is safe."
2. "Okay, that is actually getting quite good. How dare they steal my data. This isn't legal!"
3. (And then some get to the point of): "Wow that's so cool - look what I created with this!"

I have been getting a front-seat to the drama after being brigaded on Twitter for days by artists for posting about Midjourney. And I totally get it - imagine you've spent years perfecting your drawing skills, you're lucky enough to be paid for what you love to do, and then some nerd tweets about how cool a technology is, a technology that's seemingly automating your entire livelihood. I'd be upset too!

Of course this kind of backlash isn't new. Throughout human history new technologies come along, automate some jobs, and people get annoyed. Indeed the word saboteur derives from the French for sabot, a type of clog, after a group of farm-workers threw their shoes into the very machinery that replaced them.

But is AI any different? Certainly the power and rate of progress of these tools seems to be unprecedented. But I'd argue that in these specific areas, code and image generation, the initial backlash will be followed by adoption. You must remember that, not long ago, Photoshop was the pariah 'destroying' art. And before that it was photography. It seems clear that, for the moment, an illustrator is still needed. And we are already already seeing examples of artists using these tools to augment their work.

Which jobs are going to be transformed by AI next? I think large-language-models (LLVMs), like GPT3, will automate a lot of sales, marketing, and support. Most of the technology is already here - we just need to refine it [3].

Driverless cars, while infamous for being 'just around the corner', are in fact just around the corner. Cruise launched their Uber competitor in San Francisco late last year and today you can watch them successfully navigate the craziness going on in that city without a driver at the wheel. And Tesla's new self-driving beta is apparently extremely good. It's only a matter of time before this is mainstream.

What I worry about most is professions where there is no clear ‘upgrade path’. A case in point are truck drivers, who have one of the most common occupations in the States (to the tune of 3.5 million). Driverless trucks are not going to augment their drivers, they’re going to replace them. What will they do?

In the near term my prediction is that we'll have a large backlash against AI. In fact, if I were an AI researcher right now, I'd be quite paranoid about this to the point of scrubbing my information off the web.

And, because we humans have a need to personify things, any backlash will probably be against a specific company, or someone charismatic like Elon Musk. They'll be hauled in front of congress and given a stern talking to. And people will call for regulation.

But this kind of technology is extremely hard to regulate, and not just because our representatives are email-printing-geriatrics. The obvious place to start is with the training datasets which, for the most part, are private scrapes of the web. But how is regulating that going to work?

We could ban crawling the web, but that would just hand an AI monopoly to existing players like Google. We could try compensating people when their data is used? But that seems not only technically infeasible but economically infeasible as well. A single person's data is essentially worthless, only in aggregate is it worth anything (ask Facebook).

Or perhaps we could try to regulate it from a copyright perspective. But derivative art has long been enshrined in copyright law as otherwise we’d have no art (it’s all derivative). And really the regulation should be aimed at the future, the risks associated with an AGI - but if a genius AI researcher can't figure out AI safety, how is a congressman going to?

Regardless, if one country regulates or bans these AIs the work will still happen — other countries will simply take the lead.

What's next?

Shit’s about to get weird.

The technology behind these image-gen algs will be applied to video next, and then music. We are going to see an explosion of personalization and creativity. With after-shocks of disinformation and fakery.

Large sectors will be automated, like fast-food workers, receptionists, security guards, waiters, therapists, and more. Hotels will be completely self-service; first the front-desk will go, and then the cleaning. Union backlashes and rising wages will only further incentivize companies to automate.

Spotify's Discover Weekly is not just going to be curated for you, but generated for you. Your kids are going to have AI friends with customizable parental controls (like 'teach them a second language'). Emails and texts will start auto-completing to the point where you'll just hit send. Avatars will be normalized - reality will start blending with fiction.

We are going to get some incredible Siri-like assistants that will organize a lot of our personal lives. Making reservations, booking flights, researching information - all of these things can be delegated to your digital-friend. I think Google will have an incentive here to release what they've been working on since AI is Apple's achilles heel–it's one of the few things that could persuade iPhone users to switch.

What we're working towards

Why are we working on this? As technologists, we need to come up with a really good answer. A mob of illustrators are already sharpening their Apple Pencils.

The truth is, we already have a really good answer. Unfortunately it just sounds so insane we're embarrassed to talk about it.

I'm talking about the singularity. The point at which the AI starts improving itself, creating exponential intelligence.

For most AI researchers, it seems it's not a case of if but when we reach the singularity. Some believe it's hundreds of years away. Some say by mid-century. But increasingly my friends in the field have brought down their timelines to the point where some believe it'll happen within a decade.

Regardless of the timeline, if artificial general intelligence (AGI) is possible, the implications are profound.

The thinking goes like this: all problems are due to lack of knowledge. Knowledge is created through creativity refined with criticism. So, once our AIs are creative, they can solve all our problems.

By all of our problems, I'm talking about total abundance, unlimited food, diseases cured, pollution solved, asteroids diverted, galaxies explored, planets terraformed... and of course, immortality.

No wonder we don't talk about it — it sounds nuts. (But then so does any sufficiently advanced technology.)

In April Google announced PaLM, the largest language model in the world coming in at a staggering 540 billion parameters. This model can understand humor and sarcasm, distinguish cause and effect, and is approaching the comprehension of a nine-year-old.

Here's an important part of their announcement:

Interestingly, we note that PaLM’s performance as a function of scale follows a log-linear behavior similar to prior models, suggesting that performance improvements from scale have not yet plateaued.

In other words, the bigger the model the better it performs – with seemingly no cap in sight. OpenAI refers to this phenomenon as a Scaling Law [1]. They released a paper showing that, since 2012, the amount of compute used in the largest AI training runs has been increasing exponentially with a 3.4-month doubling time. Later they released another paper showing performance scaling with model size [2].

If these researchers are right, and AGI is just a function of compute, then we are well on our way.

Let’s just assume for a second that building an AGI is possible - should we do it?

The risks associated with building an AGI are largely unknown and could be catastrophic. But we do know the risks of not developing it— the suffering humans experience everyday. Famine, car accidents, cancer, natural disasters. And all with a backdrop of an existing existential risk: nuclear war.

And there's another risk, one that I am personally more afraid of, of an authoritarian nation like China reaching it first. We don’t know if an AGI can be controlled, but if it can be then it would surely pose a formidable weapon.

We are now in this uncanny valley between really good AI (which is taking people’s jobs) and the singularity (where nobody needs a job).

AI automation will widen class divides. Some people will be paid a lot more for what they do, and some will be made redundant. It’s going to be a rocky transition. So my personal belief is that we should do everything we can to bring the singularity forward and shorten this period.

In the near-term I think we are going to have to get much better at talking about why we’re building these AIs and what we’re trying to work towards (however insane it might sound).

If we are automating entire industries, we at least owe them an answer.

[1] - Worth noting there are some subsequent papers from DeepMind that show there's a lot more optimization to be made without scaling.

[2] For Rich Sulton, a distinguished researcher at DeepMind, this was his biggest learning.

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.

[3] - And by refine I mean we need to deal with the abuse problem. The issue with LLMs is that they can go 'off-message', like when a Microsoft Twitter bot started endorsing the Third Reich. Currently the only solution is filtering the inputs and outputs of these models - not a great one. But we'll figure it out.

[4] - Google also has an algorithm (of course they do) called Imagen. But they're not letting anyone outside the company use it. This makes sense— they have no incentive to release it. They make lots of money from their core business and somehow have managed to avoid a lot of the scrutiny directed towards other tech companies–there's no wish to upset the apple cart. OpenAI, on the other hand, has no such money printing machine. They are very much incentivized to productize their research.