Most AI startups are doomed
weightythoughts.com
weightythoughts.com
Well, this was only mostly true. With search engines, there was a "winner-take-all" effect. Yes, many companies could build search engines, but Google was just a little bit better. Once one of the search engines is a little bit better, why would you use anything else?
Eventually, Google figured out how to create a real moat, by using click data to improve search result ranking. Even though Microsoft is willing to spend billions of dollars on Bing, they don't have access to Google's user data, and aren't quite able to match Google's search quality.
I believe that many AI startups will have a similar "data moat". If you are the first AI company to get a significant amount of users, you may be able to learn from their behavior to improve the product. If you can do this, you'll have an advantage that competitors won't be able to easily copy.
So just make something people want, gather data on what your users are doing, and use that data to make your product better. If you do that right, you'll keep growing, and you'll be able to describe this simple strategy as a "proprietary data advantage" to give your slides more buzzwords if you need them.
Anyway, I think your point here is interesting and was kind of the idea behind a lot of the "gather lots of data" startups. A lot of those failed in part because the frontier of AI is moving pretty quickly. You need a lot less data to do interesting thing today than you did not that long ago. Because we've thrown more and more data at more and more compute, I think people don't appreciate how much we've truly progressed algorithmically. You need an order of magnitude less data to do the same thing for each "generation" of AI.
That frontier cuts against the ability to build a moat on user-generated data, so long as it's readily available or somewhat replicable. Your competitor is naturally going to have a cheaper time getting into market than you if they wait longer to do so.
However, this definitely does stand if your area truly is obscure (e.g. specific industry), annoying to gather data in (e.g. certain healthcare applications), or actually proprietary (e.g. your own device data with a different modality).
Not putting words into your mouth that you aren't saying the latter here—just making a distinction since it's easy to imagine any data being a moat, which is a common mistake I see.
Is search winner takes all? Or is it advertising?
The answer (more logically) should kind of be neither. Advertising obviously has a lot of channels, and even though Google has both Adwords and its display advertising network, it doesn't follow those really need to be the same provider... at least outside of more data to do more targeted ads. But, again, advertising dollars will follow platforms that price for ROI.
Better targeting mainly adds to the amount that Google and Facebook can charge for their ads and still have companies pay for them. It doesn't really add to their dominance directly (I say directly since, obviously, more money can buy more R&D/employees/regulatory capture/acquiring competitors/just-paying-for-dominance like with Google paying Apple. But that's all indirect).
The argument I generally hear from certain microeconomists is that they still expect there to be value in niches, given Google's highly general nature. If you're looking for super specific topics, it often doesn't perform extremely well. You'd find it valuable to go to a resource tailored for your area.
Anecdotal, but I've personally found it to be true—for specific hobbies, or for more "real" reviews, I search reddit. Except I use Google to search reddit, since reddit's search sucks, but still. Amazon or Etsy or whatever can be considered "search engines" for highly specific topics (purchases, and purchases of a specific type of product) and they do have success there too, but Google is still often the front-page to get people to those sites.
Maybe it's just that Google is just a default "front-page" and enough tech-non-savvy people just use it to get to where they want to go (e.g. the classic "type Facebook into Google to get to Facebook") that it sticks. That's maybe the most compelling reason I've heard, but it is also somewhat unsatisfying as well (as well as precarious if the defaults ever change—but maybe that's true!).
I think that is true, but only within a very narrow band of topics that are broad enough to require a search functionality (lexusnexus, webmd, arxiv, etc). I think most topic niches that I would be interested in are more often served by niche publications (ie I wouldn't need/want a search engine geared toward photography, I would mostly go to specific publications and sites that I trust).
The power of defaults, mostly.
The average user experience of picking up an internet-connected device has been very intentionally cultivated by Google. Whether you're in your browser or on your phone, Google's spent a lot of money building up Chrome as a browser ecosystem, Android on mobile, and paying off Apple on iPhones and competing browser vendors like Firefox, to guarantee that, whenever possible, Google is always the default search engine. The only non-Google default will typically be on Edge, which only has about 5-6% penetration. Since Google historically has always been the best search engine in the space, does not explicitly charge its users money, and (at least for average users) is really good at surfacing what they're looking for, most users feel no need to look elsewhere for a search engine because the default just works, switching would demand an effort, and Google is what they'd want anyways. The moat isn't big, but Google has put a ton of work into ensuing that any competing search engine requires an intentional and active choice of users to seek you out while they're worse.
At least until the recent AI play by Bing, this tiny moat was always sufficient, because if you start from scratch on search, you're essentially guaranteed to be worse, and all other 'serious' offerings under the hood were weak alternatives: essentially one of "Bing search API wrappers" (worse results), "nation-state-actor search engines" (for most users, worse results), or "Google, but with some cursory privacy measures, a subscription fee, or filtration features" (which wasn't something most users care about).
Recent chat AI represents a competing alternative to doing a search in the first place, which jeopardizes the "we have essentially all defaults and users can't be assed to switch to a worse search" barrier to entry that Google historically relies on, which is ringing alarm bells for them.
I was nodding along until I got to this. Google had, past tense, excellent search results. Now they are at best, a solid mid-tier search product that I often find myself abandoning in favor of either DDG or even Bing on occasion.
IME, Bing shines in one particular area, which is location/near me type searches. It's peerless in this space in particular and IMO they should be leaning into it more in their marketing. Google can get me the best seller of gizmos on the Internet, but if I want to go to a store and get something that day, Bing is better at that.
Meanwhile Google is steadily trending downwards in very nerdy niche searches, which is a shame because it used to be quite good at them. You specify terms in your quotes or block with minuses, but these are treated as "suggestions" now, that are overridden if Google's mystery algorithm decides that you don't actually know what you need despite directly expressing it to the bloody thing, especially if their "correction" means they can direct you to buy something even if you don't want to actually buy anything.
And, even when you want to buy things... perfect example: I wanted a small set of drawers for a particularly tight alcove in my desk that's otherwise wasted space. I spent some time on Amazon for awhile but amazon's search is even worse than google's, so I googled "closet drawer cabinet -fabric" and the -fabric bit is quite important because I was getting frustrated getting page after page of hits on Amazon that were shitty little fabric drawer setups. I wanted shitty particle board, thank you very much. And Google, in it's infinite brilliance, returned, I shit you not, a full page of shopping advertisements that were all fabric drawers.
Google is still my first go, out of habit more than anything at this point, but increasingly I find their search tool lacking and I know I'm far from alone in that.
Having a hyper profitable business that you dominate can provide a cash moat: the ability to crush your competition by outspending them.
You can do that in lots of ways: lawyers (Microsoft was nearly sued into oblivion early on), advertising/marketing/brand building, creating a talent roach motel (they go in, they never leave) just to deprive your competition of the best people (by paying way above what your competition can match), paying for positioning in the market (for example: buying shelves at retail stores for distribution), you can even afford better networking globally to be faster by spending large sums; and so on.
Being able to buy positioning to lock out the competition, by leveraging your enormous profit machine, is a type of moat.
Google (Alphabet) can get away with that spending (re shareholders) because that's the business they're in, it's core, and it's already generating hugely, so shareholders view the spending as protecting an existing critical business (maintaining a moat in this case by continuing to pay Apple etc). Microsoft can't get away with the same spending (even though they can technically afford it), because it's a prospective business (a maybe outcome) that isn't spitting off huge profits and the return on massively ramping up spending is questionable to shareholders (who will ask questions about a missing $20b in profit next year).
Does Google have a quality / performance moat with their search product? Even if they do, given the ~$100 billion in profit at risk (for that division, it subsidizes a lot of the rest of Alphabet), it's not a question they want to find out the answer to necessarily. Instead they can spend $20 billion and not have to find out if a competitor could take them down.
Apple is now secure enough in their standing for mobile that they wouldn't be afraid to switch to Bing under the right circumstances[0][1]. Ironically I'm reminded of the days where Intel would pay Dell to not use AMD (even though they weren't always better) and even Intel was aware it was a devil's bargain and could let them get bullied.
This does all lead to an interesting question; What are the value props of all of these deals? If their product is so good, why do they need to spend the money on the default slot? And what do they know is 'missing' that makes it worth them to pay third parties to convince them?
[0] - Whether those circumstances are easily achievable is out of scope of this theorycraft.
[1] - Apple probably could buy a search engine, but they are smart enough not to for lots of reasons. (mostly from the aspects of 'attracting regulatory attention' and 'can they keep improving it')
I think it remains to be seen if AI is one of those industries that benefits from network effects or not.
A related question is: Are AI models platforms, or are they applications? If they're platforms, they'll benefit from more users and more data. If they're applications, there will be very different market economics in play.
IMHO, they're applications.
IMHO, it's hazy.
I finally built a new PC [0] with the intention of actually playing with AI stuff instead of just shitposting about the dumb things AI Chatbots do.
As another commenter responded, Stable Diffusion strikes me as a platform.
I feel like OpenAI is trying to position themselves as a 'platform', and ChatGPT is like versions of windows...almost [1]
LLaMA is amazingly ambiguous; frankly at the moment it is, to me[2], at least the easiest to 'hack'; I will say that LLaMA models feel more like 'applications' but there is still an overall 'platform'.
> I think it remains to be seen if AI is one of those industries that benefits from network effects or not.
I think it can. Aside from the aforementioned Stable Diffusion (and also speaking from experience with it,) having a 'cookbook' is handy to get started. Network effects are huge in that regard.
[0] - 10 years since my last fresh build, 5+ years since my last true upgrade.
[1] - It is worth noting people complaining about things 'breaking' when new models get released
This is revisionism -- Google was far superior to any competing search engine long before Microsoft embarked on its search engine adventures. Google was hand-coding heuristics well into the Bing era. It wasn't until Amit Singhal left Google and search that they pivoted to more machine learning techniques that could use the click data effectively.
This corresponded with the beginning of Google's long decline in search quality, buffeted mostly by the fact that their on-page quality systems were extremely sophisticated at cutting out spam and SEO. The detection was far from perfect, but so many miles ahead of the competitors that their competitors kept unearthing spam that Google had long since excluded from its index but whose fossilized remains still polluted the web.
The moat that Google had was just that they were really good at search quality and PageRank was only a small part of that. In other words, no moat at all, just a better product.
What happened to building good products?
That's a fairly effective barrier to switching for many people, as they won't want to lose their email address now, or need to go through every site they ever signed up with and change it.
For example, once upon a time building a superior web browser was hard, because the moat you had to cross was that Microsoft would always prefer their own browser, so Windows created a structural barrier to adoption of new browsers. So the amount by which you had to be better than Internet Explorer back in the day was critical, because if you were just a little better you would never get enough adoption. You had to be much better or convince Microsoft to lower the moat.
Bringing up a new stock exchange was another protected industry -- the incumbents had built a huge regulatory apparatus around themselves which meant that to compete with the incumbents, you first had to comply with a lot of regulations which were themselves molded after what the incumbents were doing. Exchanges like BATS took a long time to clear that hurdle, and many exchanges never made it there, even with superior execution technologies, and were consigned to be dark pools instead of lit markets.
Theoretically social networks have had the moat of the "network effect" of having a large user base -- Facebook is useless because all your friends are on MySpace so there's no reason to switch, and MySpace is useless because all your friends are on Friendster. The implications here are that incumbents protected by moats are fragile; once the moat is crossed they can no longer compete because they've languished, so get eaten alive (as the two examples here or say traditional taxi services).
Then you're one of today's Lucky 10,000! The idea of an "economic moat" was popularized by Warren Buffet, starting in 1986. https://xkcd.com/1053/
It’s a decade-old term [1] describing an older concept [2].
> What happened to building good products?
This is necessary, but insufficient on its own. Big “M” marketing is important. (Betamax vs VHS is the canonical example.)
[1] https://en.m.wikipedia.org/wiki/Economic_moat
[2] https://en.m.wikipedia.org/wiki/Porter%27s_five_forces_analy...
The algorithm was hand-coded, sure. You don't need machine learning in order to use click data. You just need a few people to have searched for that particular query before. When someone clicks on a result and stays there a while, it's a "long click" and you boost that search result for that query.
Google was already so far ahead in 2000 that everyone who actually cared used it -- although some people still had Altavista bookmarked and used that because they did not know any better. (The internet-using community was much smaller and more knowledgeable back then; Laptops were expensive and heavy and were used mostly by traveling businessmen, the majority had dial-up internet and AOL was the largest ISP. Essentially all internet access was through desktops running Windows 95 or Windows 98. And Google was already head and shoulders above the rest)
Yes, click data helps but really only at the head, where Google already excelled. On the tail the increasing dependence on click data (and later on machine learning and an increased focus on user-specific results) has steadily made Google worse and worse, and Google's supremacy has really only lasted this long because of their page quality measures, which is still a huge problem even now on Bing.
I don't recall people saying anything of the sort.
Tech and software have always been a commodity. Twitter is barely more than a CRUD. You always had to build your moat, i.e. network effect or data.
The only difference is whereas we used to do "tech" with "algorithms" now replace that word with "AI", and it works a lot better. Seriously, replace all instances of "AI" with "algorithms" in this article and it could've been written 20 years ago.
IMO very empty virtue signaling article.
I do go somewhat beyond that in pointing out exactly why most of these startups don't have defensibility.
Perhaps for some people it doesn't need to be said, but back when I wrote this... and now... the market seems to suggest that it isn't that obvious.
He's applying a standard analytical lens to AI startups, e.g. looking for their moats through finding differentiators in economics, data, scalability, etc. He finds that "doing AI" is not a stable enough differentiator to compel him as a VC. He then lays out his reasons. There are plenty of startups selling themselves on their AI platform and/or acumen, so it's rather automatically relevant to a VC at least.
This is a better thesis than the article's. Uber for X / AI for Y is not a valid pitch.
But the article goes further. It surmises the only two valid moats are a capital advantage (compute) or intellectual property (data). These are the most trivially-verifiable moats for a third party. Which makes sense for a VC to prioritise them. But they're far from the dominant mode of differentiation.
Plenty of "AI" start-ups will do well because they found a niche, had the right team to sell to it, and developed quickly enough to keep customers hooked. They won't win because of AI per se. But they won't lose for lack of access to more compute or special data either.
The author seems to be misguided - no one is really suggesting AI is a way to win. It's a value prop thing, not a competitive thing. SaaS was a value prop thing, not a competitive thing. Mobile was a value prop thing, not a competitive thing. WWW was a value prop thing, not a competitive thing.
However, many VCs aren't looking for a moat when they invest in Saas, they're looking for a good product with good founders and a good team.
IF you're only looking for moats, you're going to be a bad VC.
For AI stuff, is there anything like that? Why do I care if it's ChatGPT or some other AI writing my paper or my code or whatever else people do with these. The AI is (at least somewhat) fungible.
I have no idea how it relates to this discussion.
Current ‘AI’ itself virtually is (built on) a single example of an an algorithm, which is why on the surface there seems to be far less scope for differentiation. There’s room for genuinely new architectures and techniques, but that’s not what most of these ‘AI startups’ are offering (even if they pretend otherwise).
The important thing to remember is, you don't just get one roll at it. You can try as many times as you have time to do so. Most of the wildly successful people I know were wildly unsuccessful before they were wildly successful. The ones that hit it off the hop had more money up front to brute force things into success which ultimately works out to them just being able to finance their failures, not that they were without failure because of their starting position.
But yeah, this article is basically just pulp.
Good execution remains differentiated. It just requires continuous iteration, evolution and improvement.
If you build an MVP over a weekend and then pivot 100% of your efforts to fundraising and marketing, as has been the trend over the past decade, yes, you're screwed. You're building dollar apps for another App Store.
Most of the arguments the author levels would have worked against the first waves of computerization, digitisation and the emergence of the Internet, in some cases more powerfully. Yet the prediction didn't hold. Capex and IP weren't sole, or even strong, predictors of new-entrant success. For Exhibit A to the first part, see Softbank.
Levels is a good example of this.
While we're sitting here waxing poetic about moat or no moat and what to do with these AI things, he's made some 7 figures in cold hard cash revenue from building and shipping things people want.
Not a problem if you are an individual with no intent to scale, but a big problem if you are investor looking to invest 8 figures.
Once you have a few thousand users giving you cold hard cash to use your service, you also have access to way _way_ better product development and marketing information. Not to mention a lot of data you can use to fine-tune your AI in ways that a competitor starting from scratch couldn't dream to replicate.
This is a big part of why you see all these BigTech companies adding AI features. Their existing user-base is the moat.
But with the underlying basis for everything being open source, everything you learned with "continuous iteration, evolution and improvement" can be copied relatively easily.
The models may be open (or available to anyone who can pay). But so was TCP/IP and the App Store.
Nobody is going to win on the tech alone because nobody has ever won on the tech alone. If your founding team has the relationships to get you ensconced at a beachhead of customers, and your technical team can move fast enough that inertia keeps them put, you have a solid shot at denying that space to other entrants.
This one AI will be able to do everything you ask it, or give you any interface you desire, should you need one - even creating a real world-like interface inside 3d/vr systems.
It's kinda depressing - the first to succeed in AI (assuming they don't kill us all), will become the top company in the world (if not the only company we need at all, because nobody can compete at all, by the time the next competitor gets to AGI - they'll be 50 generations improved, and it's just a runaway loop of self-improvement.
I think people who know how to code and code for their day job really underestimate how hard it is to build these things. Even the weekend projects. A ton of these "weekend" projects, took a weekend to build, plus years of learning and research into the best most efficient ways of building those kinds of apps.
Building a startup is completely different from your faang / unicorn software engineer dayjob. Where everything is perfectly and comfortable setup for you. There is a team dedicated to making sure your code is deployed every day. The test harnesses are already built. You have dedicated designers telling you exactly how everything should look. It's all easy.
This is like the twitter clone effect. It's a cliche at this point, the casual "I could build twitter in a weekend". Why aren't there a million reddit clones? Why aren't there a million instagram clones? Why aren't there a million notion / canva / figma clones?
If it were that easy to replicate these things they would be out there.
Ok, but why aren't there a thousand "made in a weekend" clones that at the very least function but have no users?
> Many of today's 'AI startups' are essentially just a landing page with a few hundred lines of code generated by GPT-4 to connect with the OpenAI API
This is a strawman. Most of the AI startups that have raised significant money are not that.
Because you are using social network to connect/interact with other people? If I'm making a simple AI tool powered by LLM then I don't need other users to make it useful for you.
> This is a strawman. Most of the AI startups that have raised significant money are not that.
It's not a straw man argument if adding a condition to your statement is necessary to refute it. In fact, having to do so is essentially the definition of a straw man.
There are.
When Reddit switched from Lisp to Python, a number of Lisp users had negative things to say about that, and made their own clones over short time intervals to show just how superior Lisp is. There are a bunch of open source clones now: https://github.com/topics/reddit-clone
Twitter is even more trivial to clone, and this has more results: https://github.com/topics/twitter-clone
I'm not going to track down all the attempts to commercialize clones of either that shut down after a month, but they're out there.
You know what powerful competitors have a habit of doing? The thing that keeps them powerful? Buying those with a head start.
If you go for their core business, sure. Otherwise, this is cartoonish.
If you go for a full frontal attack on their core business, you'll probably get crushed. Short of that... working at a fang will make anyone not very scared of them.
Never forget that "powerful competitors" are slow. Very, very slow, even past 200 employees. Meetings and arguments increase the latency of delivering new products and services. Incentives start to be misaligned that make it difficult to continue delivering at the same quality ( why should I put in 2X to build 100X value, when I'm only getting 0.02%? ) Worry more about the startups that start alongside you.
I do, though, believe the author missed a large class of AI startups that I think will likely succeed in the "Wait, so what IS defensible?" section: startups that focus like a laser on very specific, semi-niche workflows where things like UX and compliance are critical. My best example of this so far is Harvey.ai, whose tagline is "Generative AI for Elite Law Firms":
1. First, elite law firms have lots of money to spend, and they'll spend it if they see an ROI.
2. There are plenty of Web 2.0 startups who won primarily because of first-mover advantage. I mean, Docusign wasn't exactly amazing, world-changing technology, but they became synonymous with "legal signatures over the internet" such that they became the default for this use case.
3. Obviously something like "generative AI for elite law firms" has tons of compliance concerns around it. If Harvey.AI can address that, it's a huge wn. As another example, I know of some big financial firms/banks that have giant committees around anything remotely label-able as "AI" because there are so many compliance concerns around AI (a system that gives you an answer with no visibility into how that answer was generated is anathema to the "everything must be auditable" mindset of the financial world). Again, Docusign is a good analogy here, because so much of their initial work was not in tech but ensuring that there was a legal framework (in many jurisdictions) that would deem internet signatures valid.
My overall point is that a UX that is highly tailored to specific, profitable use cases can still win out.
That's a terrible tagline. Why would an elite law firm want generative AI?
FWIW, their homepage is worse, it contains the single phrase "Unprecedented legal AI" and a waitlist signup button, nothing else. (Which is also unintentionally funny, because lawyers care a lot about precedents!)
If you're writing that I can only suppose you're not a lawyer. Tons and tons of legal documentation is a lot of "configurable boilerplate", with 10-20% specifics thrown in. There is loads generative AI can do to speed up the workflow of lawyers.
And no, not all lawyers are complete idiots like the ones that posted fake citations from ChatGPT into a legal briefing without vetting it first.
Debating if I want to respond to this, because there is fistfuls of cash right now in software consulting for this sort of work. Boring CRUDs and API integrations make a lot of the world go round (quietly).
The difficult part (as it has always been) is identifying these.
There are clearly defensible aspects for ai startups. Specifically I think these are: a) in-context and collaborative features (since working alone with ai through a chat box is unlikely the only way we will interact) b) gated knowledge/data (since commonly available technology can be leveraged with unique data) c) edge computing and offline usecases won't be the center piece for many classical companies and therefore can be very well exploited.
I wrote up a framework to assess LLM powered Startups/Ideas here: https://assistedeverything.substack.com/p/the-three-hills-mo...
Agreed on (b) - I think this is anyone's best shot at a moat.
Curious to see how (c) evolves. It's unclear to me whether the future of these things are running locally or whether we'll all continue hitting remote APIs
Think of the difference of using a single-user application to e.g. make mockups for websites or a collaborative environment like figma, in which you happen to also be able to have AI collaborate with you. Very different usecases and solving collaboration workflows, etc. is non trivial.
I guess for (c) both things will exist. Local will be done for 2 reasons: - data sovereignty (e.g. companies wanting to have applications that are purely trained/fine tuned on their own data; but that improvement is not shared) - privacy (anything from an AI having access to all your email and calendar up to having intimate "friendships" with AI)
Interesting hill analogy—I do broadly agree with the areas.
I dunno man, levelsio (Pieter) and Danny his friend are making millions as solopreneurs taking existing models, some light training and adding beautiful frontends to it.
You’re right they’re not VC scale but this is what excites me.
AI for indiehackers is a massive multiplier.
One person million dollars net after tax / year is a phenomenal business.
I personally think Statups that raise millions of dollars are doomed. Long live the lean cockroach startups.
Midjourney is 100M+ with 10-ish engineers. That’s a phenomenal business. They raised 0 VC.
Just a reminder, any employee can write a document saying anything.
The ones that are dependent on VC cash and making little money against open source models or cheaper solutions are going to lose the AI race to zero.
Yes, AI startups are doomed. So what? Founders can make millions with a doomed startup.
This is especially relevant wrt startups which can’t compete on compute or research: instead they must compete on something that is more defensible: unique data, first mover adv, etc.
# the whole internet to scrape
# the largest amount of gpu compute you have ever seen
# more or less open source fundamentals
thus "ai" will become a commodity, unless you have specific non-public useful data.
‘There’s also going to be a lot of value generated that is simply captured by existing industry incumbents, using their market power and scale. That’s not a free lunch for society, but is also just how capitalism works and often still generates “surplus” (Econ speak for “good stuff”) for society.’
The first paragraph is also just how capitalism works, and it arguably works better than other systems at generating uncaptured value.