A leaked Google memo is going around and it’s worth a read. The author (AI researcher inside of Google) thinks that Google has no strategic edge in the future of AI and needs to change direction. The analysis is good, but the conclusion is wrong.
In this post:
- There are different kinds of moats – the public discussion is about a special case moat
- Moats are neither necessary nor sufficient for a successful business. If that sounds irrational that’s because it is
- Your toaster is just a live wire in a box
- Google has a gigantic AI moat
- In the end this is actually good news for Google, not bad news
Here’s what the memo says:
1) Surprise! There’s no moat: It turns out there’s not much that Google (or OpenAI, or Meta, or…) can do in AI that the open source community can’t do. Yes, Google can spend the money to build very large, very good LLMs. This seemed like a big proprietary advantage, but to our surprise the smaller, open source versions of these large models are nearly just as good. Through the sheer volume and innovation, open source dominance is inevitable.
2) If there’s no moat there’s no customer: Because alternative open source AI is free and maybe just as good, people will not pay for Google’s AI. So Google’s in trouble.
So what’s wrong with this conclusion?
You don’t need a moat to succeed. Sure, a moat the size of our-AI-is-10x-better-than-yours is a great source of value, and you’d be sad to see that moat go away. But:
1) You don’t need this kind of moat, it’s just very nice to have
2) It doesn’t matter anyway because Google has a gigantic moat
3) Google is probably in a better position than it was before open source AI took off
There are two kinds of moats
There are two kinds of moats: hard moats and soft moats. A hard moat is the size of we-can-do-what-no-one-else-can-do: a patent, a technological innovation, a trade secret, licensing agreement, monopoly rights, etc. When you have a hard moat you’ve got a product that no one else is capable of copying. Business is easy because you’ve got a guarantee that no one can do what you do.
This hard moat is what the author seems to have in mind, and this hard moat has been true in AI for a few years now. The best performance came from really large models, and only Google and a few others were capable of putting up the resources and know-how required to create them. This is a very hard moat. So hard was this moat, in fact, that until recently it felt almost obvious to me that the future of NLP was just going to become API calls to OpenAI: the latest GPT is head and shoulders above any generative model you could build, and even outperforms classification/encoder models on a lot of tasks. Plus it’s cheap. Plus it doesn’t need to be fine-tuned on training data. Plus it’s not a model you have to host and productionize. Plus it’s…
A disappearing hard moat
What’s really going on? If you accept the author’s point 1) that open source beats industry, then a fundamental premise of the last few years has been overturned. Namely:
Only a handful of companies like Google have the resources to create AI from scratch (still true), so they will fundamentally control AI as a resource and profit enormously (now in question). This is because there’s no way to create or host the same quality of AI without putting up the same amount of resources (now false – compress and host leaked/open source models).
It feels like the author thought Google had a hard moat, found out it didn’t, and declared game over. It hardly matters though because Google has an enormous soft moat.
What’s a soft moat? Let’s back up: a “moat” is an idea attributed to Warren Buffett that refers to a company’s competitive advantage. This includes things like patents (hard moat), yes, but also softer things like network effects, branding, cost advantage, and scale. Let’s call these soft moats. Soft moats can be overcome in a way hard moats categorically cannot; a competitor could never copy your IP, but they could scale up, create network effects, market aggressively, etc. Hard moats prevent competition outright, soft moats just deter it.
You don’t need a moat
Even though hard moat is nice it’s neither necessary, sufficient, nor all that common for a successful business. Warren Buffett, the originator of the moat concept, is famous for investing in moat businesses like…See’s Candies and Coca-Cola? Boxed chocolate and sugar water? Where’s the moat? Can’t almost anyone make a coke-tasting sugar drink? The point is that a majority of really big, really strong businesses do just fine without a hard moat, or even much of a soft moat.
The empirical inefficient market hypothesis
Even the softest moats, as nondurable as switching costs, inertia, sunk cost, etc. play a huge role in most of our consumer decision making. With amazing frequency, people (myself included) choose and stick with sub-optimal, more expensive products every day, all day long for no good reason. This goes on just beneath the attention threshold in all corners of our lives. We buy brand over generic, pay for subscription services we don’t use, avoid switching costs in great disproportionate to the value that switching might bring…Louis Vuitton, Netflix, AT&T continue to thrive even though you hate your phone carrier, there’s nothing good to watch, and that belt they sold you costs ten times more than it should have. The variety and volume of the marketplace is a multi-trillion dollar testament to the fact that we are not rational customers, and plenty of less-than-ingenious businesses out there are doing just fine. (I suspect that STEM-ish people, including myself and the memo author, are susceptible to forgetting this, while someone who works in marketing never ever forgets this.) You don’t need a moat to have customers.
Day one, no moat businesses
Soft moats are also typically built from nothing over time as a business grows. I mention this because it means that on day one of some of today’s very successful (soft moat) businesses there was absolutely no moat and nothing to prevent ten other people from copying the idea. Despite the enormous soft moat of network scale that Facebook enjoys today, on day one there was absolutely nothing special about it. If you were there in Facebook HQ on day one you would have been entirely correct in declaring that they had no moat. Is it therefore a bad business?
You would have been entirely correct in declaring there was no day one moat for nearly every one of today’s most successful brands of clothing, retail, travel, food, media…
I believe it’s for this reason that in this AI discussion people seem to mean “hard moat” whenever they talk about moats. And the absence of a hard moat makes people mistakenly think there’s no opportunity and no future. Just because you don’t have a hard moat doesn’t mean you a) won’t have a successful business and b) won’t build a soft moat. It’ll just be harder than if you had the guarantee of a hard moat.
Today’s day one, no moat businesses: AI wrappers
This obsession with moats has played at volume over the last few months, mostly in reference to the thousands of apps (some good, some bad) that have built a paid product that looks an awful lot like a frontend wrapper around an OpenAI API call. I feel safe in saying that some of these have precisely no moat, in the sense that I can build the same thing in an afternoon. On the other hand, I didn’t sit down and build it (even out of spite) while the other person did, so they get 100 points while you and I get 0 points even as we sit and laugh at how silly their app is. And to laugh at these businesses for being completely undefended is to laugh at all of the very many successful businesses that started out the same way. I’m aware of the survivor bias here: there are multiples more moatless businesses that went nowhere, but the point stands that they are neither necessary nor sufficient on day one.
What makes a day one, no moat business a good business?
This is its own essay topic with the lots of different answers: good founders, fast iteration, hard work, etc. I’ll add that for AI wrapper companies, the “wrapper” can be laughably thin, or it can be a really well-thought out set of features, or marketing, or that you were first…. a handy feature, or good design, or even a nice user tutorial can become a soft moat surprisingly quickly. I wouldn’t call it a thin wrapper, but ChatGPT vs. GPT-3, or InstructGPT, or GPT-3.5 is really just a chat interface with a built in prompt. The tech wasn’t wildly different, it just happened to be in a chat format that caught on (Greg Brockman talked about being skeptical / very uncertain that ChatGPT would be of any interest). The meme version is to just build what people want to use. It’s OK if there’s no moat.
Reconsider the toaster
Nathan Baschez argues something similar in this thread, and made me re-evaluate my relationship with my toaster.
I’ve taken one apart to fix before and it is, in fact, just a live wire in a box. We’ve been swindled, people, there’s no toaster moat. We’re on the edge of another startling discovery: a light bulb is just a live wire behind glass.
None of this matters much anyway because Google has a gigantic soft moat
Network effects, branding, scale, switching costs…in the early days there was clearly a special sauce in Google search, but at this point endless user data, brand, and switching costs/inertia could more or less carry the company. What would it take you to switch from gmail to another email client? How bad would gmail have to become, and how good would a competitor have to be? Or Google Maps? Google search, even, seems to be on par with other search engines (I would argue it’s been worse for a while), yet in the consumer sphere there is still a kind of deeply burrowed embarrassment about ever using Bing.
The AI Utility Company
OK, there’s a sleight of hand in my argument: the author is talking about a moat specific to AI, while I’m talking about a moat specific to Google the business with its suite of Google products. But that first moat specific to AI is also specific to a very special, hard moat model of AI that we can call the AI utility company:
Google, or OpenAI, or whoever becomes the monopoly AI utility provider that pumps raw intelligence into the digital grid, turning AI on and off like internet, electricity, gas, or water. The utility provider controls the flow of AI and takes a cut of everything that is built with it in the application layer. This company becomes absolutely gigantic.
(I can’t find it now, but I recall Sam Altman talking about this kind of future maybe a month after ChatGPT launched (as if it looks/looked like we were heading there), discussing the calculus and responsibility companies like OpenAI would have in determining how to allocate the superintelligence, and to whom, and for how much.)
If you agree with the author’s premise that open source now competes with industry, then the dream of the AI utility company is over. What now?
The memo paints a bleak picture for Google’s future. On the contrary I would argue that this is all good news for Google. If you think the memo’s claim 1) is true, then Google is in an even better position than it was before because they’re now playing an easier game with less downside.
Game 1: the upstream AI utility company
If the game is about becoming a utility company and supplying raw AI, then there’s both great opportunity and existential threat. If a competitor like OpenAI wins this game and gets a hard moat in AI (I think, given the last two years, OpenAI does beat Google in this game – ironically Bard is definitely the Bing of AI), then Google is in serious trouble. A better AI could eat into most of Google’s products, in particular search, and takes a cut of every other AI product on the market. Therefore Google needs to win this game, tie this game (diluting its present-day dominance and splitting the utilities grid with competitors), or needs a subtler outcome: the game has no winner. If there’s no winner and no one can profit from supplying raw AI, then where’s the money to be made?
Game 2: the downstream AI application layer
The first game is to sell raw AI and make a cut of everything that happens downstream in the application layer, where loads of AI products are presumably making loads of money. Since that competition is over, the game moves directly into the application layer. The new game is about AI integration and application: make great AI products. For Google this game is both easier to win and less disastrous to lose. There’s no clear winner-take-all existential risk like in the first game, and given their dominant product suite and mountain of user data (all else being equal in terms of AI capability, more data should put you in the lead), who’s better positioned to win this game than Google?
This is where Google’s soft moat makes a huge difference. For you and I, this is a day one, no moat business. For Google this is a day one, no hard moat, gigantic soft moat business. Most of us use Google products every day to access humanity’s collective knowledge, conduct business, talk with friends, find food, navigate the surface of the earth, etc. They’re in the unique position (save Microsoft/OpenAI and to a lesser extent Meta, Apple, Amazon) of having all of the pieces: products, users, AI. Putting them together won’t necessarily be simple, but it’ll be a simpler game than it is for anyone else. In the end it’s a game Google should prefer to play.