Generative AI: What Companies Will Win?
This is the follow-up to yesterday’s article, Generative AI: Everything You Need to Know.
As you can tell, some industries are about to disappear.
Marketing copy is about to disappear.
Stock image companies are about to disappear.
Corporate photoshoot photographers are about to disappear.
And rich-world customer-service companies.
The race is on.
What industries will get decimated? Which ones will just be transformed? Which new ones will appear? If you are considering working or investing in this field, what should you be looking for? What are the pitfalls?
What’s the Winning Formula?
I went through something similar 12 years ago. I landed my first tech job in Silicon Valley in 2010, during the explosion of social media and mobile. There was a frenzy of apps on both platforms. Companies like Zynga, which eventually sold for $13B, were created at that time, on the back of Facebook’s social graph. Others like Uber appeared on the back of the mobile platforms.
What people don’t remember is that many companies exploded and disappeared. I joined a company called RockYou when it was worth $500M, but a few years later, it went extinct. There were dozens of stories like that one. It’s important to understand what happened there, to draw the right lessons.
First, the early winners were companies that went very fast. They didn’t wait for the industry to be mature. They jumped right in as it was developing. They studied Facebook’s feed APIand the App Store’s mechanics, put stuff out in days, and raked in the downloads.
What types of apps did they make? Copies of what had already worked. There were companies making horoscopes, birthday cards, funny videos, simple games, and the like, who got to tens of millions of customers in a matter of months. Same thing on mobile. The first apps and games were copies of existing ones, like poker, scrabble, rock band, and so on.
We can see the same thing now for example with the fast independent developer @levelsio, who built avatarai.me and interiorai.com. He’s shipped more than 70 projects, only 5% succeeded, but the latest one to hit made him $100k in 10 days. Pick things that already exist to reduce the risk, ship fast, try many things, and one might succeed.
Understand Your Power
Facebook grew between 2008 and 2012 on the back of the frenzy of these social media apps. Then, they copied some apps (like birthday celebrations), and killed the rest. Because all these companies depended on the access of Facebook’s news feed to get new customers and retain the existing ones, once Facebook turned off the faucet, they all dehydrated.
We’ve seen the exact same pattern with Apple and Google, who tax their apps, copy the ones that make sense, and shut down the ones they dislike.
In AI, the equivalents of Facebook’s feed, the App Store, or Google Play are OpenAI’s models for GPT-3, DALL-E, or Whisper, or other models like Google’s PaLM or LaMDA. You can’t build a business on top of a quality AI. Best case, others will copy you. Worst case, the model owners will shut you down.
It’s one of the reasons why it’s so good for the industry that there are alternative models like Stable Diffusion and Midjourney—especially Stable Diffusion, since it’s open source.
But even without the risk of being taken down, you still have the risk of being copied as soon as your product is successful. So you should only invest in businesses that can be defensible.
As we saw, there are dime-a-dozen companies that help you write better. I would not get into that business—unless you’re the company with the most money and the best engineers. They can build many features faster than the competition, which will attract customers who don’t want to use ten services for one task. The more of them they get, the more PR they’ll get, more word of mouth, and eventually they will beat smaller competitors.
All of this means that B2Bmight be easier to tackle than B2C: Customers are more knowledgeable, they’re richer, and very sticky. If you have a sales or marketing advantage, you can win a share of the market, which other companies will look at to feel comfortable buying your products, it will become easier to grow, and you will be able to beat the competition. This is especially true if you’re amongst the first to bring AI to more obscure industries.
More generally, all of this points to a stark fact: For most AI companies, an advantage in the go-to-market plan (GTM) will account for the majority of their success.
Another way to approach this is following what Replika did. Creating avatars that learn from you and start mimicking what you say and how you say it is defensible: You don’t want to move to another service and start training your avatar from scratch. The more an AI learns from the customer, the better.
That creates retention. Good. But not good enough. Ideally, you build a business that has retention and user acquisition baked in. That’s why marketplaces have grown so fast in the Internet era: The more customers, the more valuable the marketplace, so the more people come, and the more they invite others to come to the marketplace so they can also benefit and contribute. They have network effects.
What is the equivalent for AI?
Two ways come to mind. The first is the same: Human network effects. For example, if you create AI collaboration tools, they will be stickier and more viral.
Another example might be marketplaces for AI-enhanced services. Maybe you would be ok paying an artist for one hour to quickly iterate on some Midjourney images: Prompt engineering (finding the prompts that create amazing images that respect your vision) is hard. Which is why some of the first companies to appear in the space are prompt engineering marketplaces, where you can see other people’s images and prompts.
The trouble with marketplaces is that the network effects come from humans being on the supply and demand sides. But the AI either replaces humans, in which case there are no network effects, or they enhance them, in which case the marketplace will likely be the same as before. So you’d need to create marketplaces specific to AI—again, one of the reasons why prompt engineering marketplaces appeared.
The perfect company would be one where there are network effects linked to the AI itself. It’s one of the reasons why companies like Stability AI (behind Stable Diffusion) raised $100M at a $1B valuation: The more customers use Stable Diffusion, the more the company learns and improves its algorithms, and the more they learn what use cases customers have, and solve for them. That’s like Google’s advantage with all the queries in the world. Hard to beat them.
So the billion dollar question becomes: What companies built with AI at the core can get better the more people use them, in a way that you might want to invite other people?
I’ll give you an example: a social network where you interact on an equal footing with AI characters.
Today, every network adds a certain type of social value:
The value you get from Twitter is the ideas of the entire network.
For Tiktok, it’s the video creativity of the network.
In Instagram, the same, plus pictures.
In Snapchat, it’s a closer group of friends.
In Facebook or LinkedIn, your real-life social and professional networks.
Maybe there’s room for more? For example, a network where you can create new AI characters for a small chat group? You could define what each of these characters is like, and chat with them, without fear of being judged. Maybe other humans can befriend those same characters, or join your chat, making it livelier.
I’ve had another idea for an AI-enhanced social network for years. I think it’s a trillion-dollar idea. It uses collaboration and AI for better decision-making. If you work in tech and are interested, let me know.
So these are the types of companies that might succeed in the future. What are some examples of industries that will be completely transformed, destroyed, or created?
Talk with Your Dead Grandmother
Imagine you could have a true conversation with your loved partner who isn’t with you anymore. Your mother, your grandpa…
Humans have had this dream for millennia: Talk with the dead.
So far, we haven’t been able to make it happen. Until now.
Many companies have tried. You have Eterni.me, from MIT grads, but it never launched. HereAfter did launch, but it sounds more like a way to access memories than talking with a deceased loved one. Re;memory is the same, this time focused more on recorded video than chat.
The founder of Replika is famous for creating a digital chatbot of her dead friend. This spawned her company, but she moved away from digital memory, and created instead your digital friend, who learns from you as you speak with it.
None of them work well. But this shouldn’t be the case.
Most of these apps were started before models like GPT-3, DALL-E, or Whisper appeared. It might be hard for them to throw away all the progress they’ve made, but they should. They can’t compete with the progress of companies like Google, OpenAI, or Meta.
Instead, they should focus on putting all the pieces together:
A great chatbot that can be trained on existing content from the target person, along with the conversation itself, and anything else happening in the world.
A great tool to create a 3D version of a person, ideally based on just a handful of pictures.
A great tool to synthesize the voice of a person, ideally from just a few recordings.
A user experience that can put all three together well. For example, whatever these companies do, the first few interactions with the AI must be mind-blowing.
Talk with Obama
Once this works, this can be extended to other people. For example, for a small fee, you could talk with AIs trained on famous people. Today, this is impossible. And if a famous person offers this service, you’re not really talking with them, but with their assistant. Trained AIs will soon be better at mimicking people than their assistants ever could.
How much would you pay to have a conversation with Obama? With JFK? Feynman? Hitler? Gandhi? Jesus? How much would you pay to see a conversation between them?
If you think this is a small industry, think again. This is the basis of the monetization of companies like Twitch or OnlyFans. People pay tens of billions to interact with famous people. But these interactions are not scalable. How does somebody talk with one million fans, one on one?
Talking about OnlyFans…
Sex is an obvious variant of this. Today, sexual content is mostly focused on recorded videos. AI could do it so much better:
You can define the actors.
You can define the acting.
You can set it up in 3D.
You can participate.
The AIs can learn what works better with you or people like you, and take it over as it learns.
I haven’t seen any company trying to do this, and my guess is that most people shun this type of work. The result is that the few that dare getting into this will have a huge advantage. That’s what you want: No competition.
Love and Friendship
This can also work as a form of love and friendship.
Many people don’t have access to love today, or have personalities that make it hard for others to stay with them. Should they be deprived of love and friendship?
Even for those who do have friends or partners, these relationships are usually hard. We’re not always as mature as we would like. We lack mental frameworks to deal with relationships perfectly. We are selfish.
AIs can be better than that.
They can be understanding, thoughtful, compassionate, wise…
Replika is the closest thing to that today. I just tried it for a few minutes and I found it very underwhelming. And yet, when you listen to the people who have formed a relationship with their Replika, the words that come to mind are friendship and love. Imagine applying a better conversational tool to it, like LaMDA, and a more realistic 3D avatar.
Siri and the like have been nearly there, but not quite: They don’t quite understand everything you say, and their ranges of responses are limited. That made them nearly unusable.
But take a Whisper model to better understand you, a LaMDA model to have a proper conversation with you, and a GPT-3 model to answer more convoluted queries, and you have a fully-formed assistant. Make it learn from you over time, and it becomes a truly personal assistant.
Instead of training them on yourself, you can train them on a specialized body of knowledge, and they will become specialized assistants. What if you want to spend two days in London with your children, but you’re not sure what to do, and are flexible on the dates? That’s two hours of research right there, but an AI could solve it in seconds.
The same is true for any other topic. Imagine loading an AI on your field of expertise, so you can have conversations with it, asking it questions.
Here’s an example. You can see my question on the top left, the summary on the left, and the papers and their summaries on the right:
Now add a lab run by robots, and imagine how the speed of research will accelerate.
One of the issues in education is that everything is dry. We ask students to imagine the concepts we tell them about. But what if they didn’t have to imagine?
This museum recorded thousands of answers from a holocaust survivor, and created a hologram of it. But what if it added a layer of AI, where it doesn’t just repeat canned answers?
This person created GoPro footage of the French Revolution.
Obviously it’s not there yet, but it will probably take little time to get really good. We’re talking years, not decades!
Convert a document into quizzes with Quizz Gecko. For example, I looked on Google for how to terraform Mars, loaded the url in here, and I got these questions:
Incredible! Although it’s not ready for prime time. As Shoni highlights, the 3rd question gives away the answer to the first question.
But what about going beyond all of this, and creating a personal tutor for every child? Personal tutoring is the most powerful educational intervention we know about. Create an AI tutor that works, and human ability will explode.
I don’t think educators will be replaced by AI any time soon: A big piece of their role is childcare, another big part is to show another human you care about them and expect them to challenge themselves. And in many cases, the AI will just be plain wrong.
But all the parts of the job that are menial, or hard to scale, will be taken over by AI. Like identifying where exactly every student is standing in their performance, or explaining to them for the 7th time and in a different way a concept they’re having a hard time grasping.
The same will be true for artists, photographers, and the like—although here, it looks like a bigger piece of their job will be automated sooner.
Corporate headshots will soon be a thing of the past. They already nearly are, with things like avatarai.me.
Family shots will also change. Today, they’re expensive. Imagine if you could take 20 pictures from each of your family members, send them to an AI, pick ideal poses and environments, and use that to create the most amazing family photoshoots you’ve ever seen.
The thing that this will not achieve is to capture a moment. But anybody who has a family with small children and has suffered one knows today family shots are not a moment to be captured—you might want to forget them. Instead, it might be much better to take pictures of an event, upload them to an AI, and let the AI craft from scratch images of poses that never happened, but truly capture the spirit of the event.
There’s an intense debate on whether artists will disappear or not. Some are understandably scared. But which ones will disappear, and which ones won’t?
From my article What NFTs Can Learn from Art and Luxury:
“The advent of photography wreaked havoc on the realistic aesthetic in painting. Painters could no longer hope to impress viewers by depicting scenes as accurately as possible, as they had strived to do for millennia. ‘In response, painters invented new genres based on new, non-representational aesthetics: impressionism, cubism, expressionism, surrealism, abstraction. Signs of handmade authenticity became more important than representational skill.’”—Culture critic from the 1920s Walter Benjamin, through Robin Hanson, Kevin Simler, The Elephant in the Brain.
In other words: The representation of reality was valuable only because it was scarce. The moment it wasn’t scarce anymore (because of photography), the representation of reality itself lost all value, and people started valuing other things. Showing that it was never the image itself that was valuable, but rather its scarcity.
Artists will not disappear. In some cases, we value that it’s a human who did something.
In all the other cases, they might disappear. If you don’t care about owning a piece of art, but rather just need some illustration, you will not want to pay a human: They are expensive and take hours for each iteration instead of seconds.
These are some ways that AI people emulators might change the world in the coming years. But we might not need fully-formed assistants to be better. We can also use AI to improve the productivity of different industries, by completely revisiting the work processes.
This week’s previous article gave you a sense of how movie-making will change. Soon, you will be able to easily:
Brainstorm ideas for scripts. Iterate on them.
Create dialogue from text. Make it more poignant, more emotional.
Create actors from scratch.
Direct their body and face. Tweak their performances until they’re perfect.
Create any environment.
Edit the environment at will, including lighting, atmosphere, colors…
Add any object at will.
Add any special effect.
Edit the vocal performance until it’s perfect: the right tone, pitch, intonation…
Create music from scratch, which will match the mood and cadence of the movie perfectly.
Create sound effects from scratch, tailor-made to the movie.
Every bullet point from this list already has a basic tool to do it. We’re not in the inventing phase anymore. We’re in the improving and deploying phase.
You will still need humans to direct it all (at least at the beginning). And in many cases, the tools will only replace the most basic tasks, or those from the least skilled professionals. But they will get better over time. We might be just a few years away from individuals being able to make movies end to end. Movies and shows might go the route of social media: Most of the posts will be crap, but a few creators who didn’t have access to the right people and money in the past will now be freed, and their creations will take over the world.
What’s true for movies is also true for videogames, and people are already doing it.
Creators are using AI for concept art.
And then plugging it directly into the game!
You can then put these assets in production, like in this game, where all assets are generated by AI.
And whole universes for your players, objects, and NPCs
Since 50% of the cost and time spent building AAA videogames are invested in art and design, videogames companies will be able to build games for half the cost, in half the time.
A big chunk of their cost is quality assurance and playtesting, and this might be automated too.
Many of these tools are not there yet. But some are close to ready. For the rest, we don’t know how fast they’ll improve in the coming months and years. But if the last few weeks are any indication, we can bet on fast improvement.
The best products are made by frustrated users: They know intimately well their own pain points, so they’re very focused on solving them well. It’s no surprise that AI for coding is so advanced, and that it will keep progressing: AI engineers prioritize improving their own workflows.
In the last article, we talked about Replit’s Ghostwriter and GitHub’s CoPilot and AI pull requests. We can expect more and more parts of engineers’ work to be automated by AI. Development might evolve into a split between all the parts that AI can make, and the few really hard pieces for which humans are still better. Those who build things AI can’t make might make even more money than today—since they’re more productive thanks to AI enhancement.
Meanwhile, those whose abilities can be replaced by AI might become more like managers, who stitch together existing tools to quickly create applications.
If you push this forward a bit further, this will meet the No Code trend and anybody will be able to build any application they want. It’s like the youtube world of videos, but for applications. Imagine you could talk to an AI to describe to it the mobile app you want. As you describe it, it gets built, and you can play with it and iterate for improvements. Instead of spending months or years building a software company, you could spend minutes. Most of the applications built this way would be crap, but a few would be amazing, built by people who today have amazing ideas but not enough access to making products. Imagine the type of amazing progress we could witness!
Making and Selling Physical Products
While software development is focused on bits, the same can be done with atoms. Today, you can’t easily study customers and understand their pain points. It’s one of the few pieces of the process of making products that you can’t enhance with AI yet. But once you understand what customers need, you can brainstorm and quickly iterate on new designs. Then, you can play with patterns, tiling, textures, and any other type of product characteristic you can imagine. You can then create experiences for customers, where they can try the products themselves.
Remember the fashion example:
This is a proof of concept. Imagine once it’s perfected!
Or interior design:
This is where product development bleeds into marketing. If customers can personalize their own shirt, that’s like a digital salesperson.
You can lean on that idea and push for AI-led marketing. Remember this?
Going from product to content for marketing is now easy.
All of this Is a glimpse of the future of marketing. We will soon be able to:
Research and brainstorm great offers.
Brainstorm brand names.
Generate brand logos.
Brainstorm and build amazing creative assets.
Automatically test creative in ad campaigns.
Precisely attribute it across channels.
Optimize ad campaigns.
Give them personal assistants to explore all of this.
What other industries will change?
What other types of companies can structurally win given this technology?
Where would you bet your time and money?
OpenGraph at the time.
Business to business. As opposed to companies who sell directly to consumers.
Your faithful editor Shoni shared a nice example about this. In her words: “Last night I watched a documentary about Jurassic park, which started out using animatronic dinosaurs, then basically threw out all their progress and started again when they saw what CGI could do. The stop-motion guy got sick and depressed, but ended up being the dinosaur movement consultant or something, so his skills were still useful in the end.”
OnlyFans works for any creator who wants to sell a direct relationship with their fans, but it’s especially famous for female creators selling virtual sex.
I am not exploring the ethical issues here, but they exist. As we get into the world of AI, one of the key questions will become: What is real, and what is AI-generated? For example, on this topic of history, what if these imagined conversations begin to be treated as the equal of witness from actual survivors? What if the AI version began to introduce subtle variations? What if the AI learned and began to invent versions that overtook the historical record? And if you have AI versions, why go back to history itself? Why not just satisfy yourself with AI history? Looking into the historical record, you make discoveries that rely on leaps of intuition and connection that AI can't make, or would never make.