.ai in your startup does not make it an AI startup.

Every few years, a new trend emerges, capturing the collective imagination. It stirs a frenzy, spawning conferences and reports, birthing startups and luring investors. Picture a scene from Pirates of the Caribbean, with pirates gathered in a dimly lit bar, swapping tales of treasure islands overflowing with riches. We live through our own version of this scene, driven by the pursuit of the next big thing. While many of these quests amount to nothing more than wild goose chases, a select few do uncover genuine treasures, sustaining this cycle of excitement and ambition. In recent months, the focus has shifted from Web3, D2C etc to AI. As a result of this everyone seems to have rebranded their startups with a “.ai” extension, signaling their alignment with the AI treasure island.

Before the ChatGPT era of AI it involved multiple PhDs to do anything in AI. Today anyone that can hook up an API to GPT such as the one from OpenAI can build an AI application. So everyone is an AI startup. With GPT everyone has a ‘Black Pearl’ to steer their voyage. .ai rebranding does not give you any leg up or advantage to show you are unique. Or that you will succeed in the voyage faster compared to others. To the eyes of the average investor there is nothing special.

In this post-ChatGPT era, the first question you should ask as an AI startup is whether you are building applications on top of LLM (large language models) or developing the LLMs infrastructure. Earlier in AI applications, everything was intertwined and custom-built, much like the days of mainframe programming when there was no concept of an operating system. Think of LLMs as like Operating Systems. Just like how there are only a few large operating system companies in the world, to quote Sam Altman “the age of Giant large language models of AI is over”. Unless you are building an Operating System focussed startup you should not build your own Operating System for your application.

Rules of the game to win for application level software and systems software companies are different.

The biggest issue founders and investors face today is the FOG that is being generated by tectonic shifts that are underway in machine learning. Gaining clarity about how these different games emerge, identifying the right people to talk to and talking the right game can eliminate half the problem around funding conversation.

In the world of funded startups, you must win the narrative before winning the race. Investors invest in business stories, not the businesses themselves. This time, the story must be slightly different, as some narratives have been repeated in the past and have disappointed. Moreover, it needs to actively overcome significant objections of the past narratives.

Scarred by experience, you don’t touch again

AI is unique among trends in that it has undergone two winters. The first winter, which took place in the 1970s and 1980s, saw a surge of government and private funding for AI research that ultimately failed to meet expectations. This led to a sharp reduction in funding. The second winter occurred in the early 2000s when many AI startups failed following the dot-com bubble burst. Collective memory and scars from this past still has a big influence on what gets funded.

Even in the last 20 years there was a peak up until 2015 and then a sharp drop. At least that is what data I pulled out seemed to indicate. Now this part was debated. A founder friend skilled in handling numbers pointed out that the figures I cited in my previous post on data regarding AI investments, particularly in recent years, seem low. He is most likely right, I didn’t work with an expert analyst to do a thorough research. Instead, I extracted the data myself from Crunchbase for each year marked as AI investments and reported the numbers as they appeared. These figures might have been over- or under-counted. Some companies incorrectly classified themselves as AI or vice versa, which can influence the data badly. Additionally, Big Data is a related field, and it is unclear whether its funding was counted within AI investments. This shifted the debate to definitions, which is a pointless exercise.

I started looking for reports such as Pitchbook, CBinsights and others to see what they say. Most reports only have the incentive to paint a rosy picture.

Getting to raw data and cleaning it up to be very objective with data is very hard exercise. Had to remind myself that I should stop torturing the data. Poor data, every time it is tortured it will confess to whatever insights you want.

Speaking to a few investors, anecdotally it helped me get convinced again that indeed a lot of investments had gone in. Not a lot of money was made out of those investments. It would be reasonable to argue that investors that invested in AI made far less than those who did in marketplaces. While it is possible to drench any category with excess investments, marketplaces like business models require far less capital to become self-sustainable and also grow on their own.

If you pitch an AI startup to investors who went through an AI winter before, you are less likely to convince them. They carry the baggage of how AI investments didn’t work out for them and may hesitate to imagine new possibilities.

Gen Z VCs, unscarred by past experiences, will imagine new possibilities. Similarly, seasoned investors who can set aside past experiences and evaluate each investment based on first principles can maintain a neutral perspective. Prioritise meeting them, skip others

Retaining any fizz is hard

The recent excitement in AI is about Generative AI. It is just the fizz. Generative means prediction or guessing. It is like those students who mug up their subjects and then vomit out. Machines are taught now with a corpus of mugging up the entire web. When they encounter a question that is out of that syllabus it turns out that they will still guess and give an answer. A confidently guessed answer. Whether it is fact or not, we are impressed by it. This is what Generative AI is doing underneath. Ability for the AI to make confident guesses that impresses. Building a generative AI application startup could be a very tricky thing. Most generative AI startups think that a student that has mugged up the entire web (GPT) can mug another chapter, for example your website or the corpus of your support queries. This is broadly called fine tuning. Yet, generative power appears to be better addressed through scaling. Any generative element you add can be rendered obsolete by the next two GPT revisions. Generative AI startups then could end up becoming like that application developer building a great app only to become obsolete in the next Mac OS upgrade. Investors worry about and founders should think about that too.

Where there is buzz, there is crowd. But the crowd that comes quickly, goes away quickly. This is why investors that look into applications look for retention curves. How many stayed back and who were they ? Strength of a good application is measured by retention, not even revenue initially. In business applications it has been referred to by its cousin term churn. Not just churn, the desirable thing is negative churn. Customers that stayed back how much more they spent. Being an AI application on top of an LLM does not change this dynamic of application success.

In business applications workflow integration is key. Integrating in existing behavior and flow would make adoption not only easy but provide anchors for retention

Smaller businesses may not be as concerned about data and privacy. They would trade off sending their data to OpenAI servers for more ease and a lower price. However, enterprises would be worried about the privacy and confidentiality of their data. Moreover, hallucination would be a deal breaker for them. Hallucination, which is an inherent property of generative AI, would not be acceptable for answers that could have legal consequences. Enterprises would require these three issues of data privacy, confidentiality and hallucination to be resolved before adopting AI applications. The resolution of these problems would make the applications mission critical. That would increase retention.

Think about retention all the time.

Double game is two times hard and window is short

Often it happens in the startup world what starts out as an application in the startup maze evolves into operating system level system software. Like Hugginface which started as building an AI tamagotchi but eventually iterated to being the Github for AI.

When that level of change happens then the game changes.

The dynamics below the LLM are different. There will be few hosting and hardware mammoths such as Google, AWS, OpenAI, DigitalOcean, and other software infrastructure players.The system software part will mirror a lot of what open source went through in enterprise in the last decade.

Databricks CEO Ali Ghodsi once said, “Winning in open source means playing and winning in two games.First cycle is like playing baseball, making that work is hitting a grand slam. In the second cycle, a company must play golf, you pick a golf club and hit a hole in one to make the business model work. “

You need two miracles and each game may take 5 to 7 years. Most of the investors’ fund cycles at 10 years can hardly support the first 7 year game. By the time the first game is over they are forced to chase game win proceeds.

First time investors and those who face career risk pressure to show fund returns cannot get in an arena that involves playing double games.

Defensibility doubts that paralyze

Moat is a favorite topic for investors. It should be. Investors seek an exclusive period of profit generation for their investments. Without this, the incentive to invest reduces. Moat was popularised by Warren Buffet and other public market value investors. However, it is difficult to think of moats in every stage of business, especially the startup stage.

There are many ways to classify moats. Jerry Nuemann has one of the most refined and exhaustive classification of moats based on their source.

Moats can be classified into state-governed, special know-how, scale-induced, and rigidity-based.

State-governed moats include licenses, legal monopolies, and patents. Special know-how moats are based on secrets, tacit knowledge, and insights. Scale-induced moats are driven by supply-side, demand-side, network effects, and discounted cost of capital. Rigidity-based moats are based on what individual, enterprise, and cultural behavior won’t change.

Technology industry resists and lobbies against most forms of regulation and state interference. The state never catches up with technology fast enough. Therefore, in the field of technology, state-governed moats are far less. There are some exceptions.

In startups, the question of scale is not relevant in the early stage. Therefore, most discussions about moats based on scale categories are not applicable. Early stage startups go through many twists and turns in their progress maze, making it difficult to predict what scale moats may exist. A network effect platform may evolve into a single application business, or a trivial application may transform into a horizontal platform.

The third type of moat that is based on a trade secret or insight is a tricky one. Sharing it would dilute its value. Therefore, confidentiality must be maintained. This makes it a non-starter conversation in the technology world investor conversations. Peter Theil’s question of ‘knowing secrets that others disagree on’ becomes relevant in this context. This disagreement becomes the protection until the startup jumps to scale and grabs another moat. However, convincing investors of this type of insight can be a difficult task.

Most people underestimate the moats that can be created in the fourth category. Asking what won’t change is a very powerful question. Humans don’t change, there is inertia in organizations, beliefs are deeply rooted in culture & religion is lindy. For example if people get used to a certain style of interaction then they don’t want to switch unless the new user interface is ten times better. Apple has shown this in the enterprise market. Once enterprise users get used to Apple’s experience, they don’t want to switch to alternatives as it is ten times better.

Those who say ‘data is a moat’ are referring to switching costs, users would not want to switch if they can’t take their data along. Connectors and integrators make that possible. That is why companies like Zapier, Mulesoft are hot acquisition targets They decrease switching costs, help increase competition

But all of this is big company conversation.

Despite that many investors ask for answers to this.

But as an early stage startup you can’t answer it in a useful way. The only competitive advantage that an early stage startup has is speed. Therefore, the only convincing answer is that you can run to one of these four sources of moats faster than others. Sadly that answer for new investors is not enough.

Depth of the moat conversations differentiates the experience of the investor.

Don’t have the AI sidecar

In life there is known, unknown and then unknowable. It is evident that what is happening in AI is in the ‘unknowable’. We now know that artificial neural networks work. To a degree that it can solve useful problems. No one knows why they work. Working in AI is like being in Star trek. “Going to places where no man has gone before”. This terrifies most people. Therefore, investing in star trek voyages is therefore only for the bravest of braves.

Rick Zeckhauser has written one of the most seminal works on the topic of investing. It is titled “Investing in known, unknown and unknowable “. He suggests that investing in the unknowable is the hardest, yet most rewarding. There are 8 maxims he presents. For investing in the unknowable he suggests having a sidecar approach. Partner with someone who has a slight edge in peeking into the future. As long as they are trustable, take their help in making bets Most investors don’t have that peek into the transformer architecture led future neural network that has caused the inflection in machine learning and AI.

Naval calls angel investing as a lottery ticket where you know one digit of the lottery ticket. Applies to all of early stage investing , seed, pre-seed, or even series A. Ask any one who has done angel investing for over twenty years and they will tell you that one digit is correlated to growth mindset of the founder. Rarely anything more can be said. In the unknowable fields you may need 2 digits instead of one. The second one must inform this slight peek into the future.

This is why investors like Andrew Ng will have an edge over others. Andrew has not only taught machine learning for years at Stanford, but has also run two companies and manages a hundred million fund. Most can’t be like that, have three careers rolled into one. In those cases look whether they have advisors who understand language models deeply.

Ideally folks like the eight authors of “Attention is all you need” research work. Interesting thing though is that they have gone on to build their own starships.

Finally

While packaging is important in everything in life, go beyond the surface level ‘look and feel’ change just calling yourself as .ai startup. Parse out on what level you are playing and how the startup maze iteration is changing your game. Above LLM or below LLM ? 1) Talk to Gen Z investors or those who are older but can reason with a beginner’s mindset. 2) Look for folks who don’t have first fund career risk and someone who doesn’t have the pressure of a 10 year window, therefore is structurally set up for picking up a double game play. 3) Defensibility questions differentiates the novices from the seasoned. Look at the quality of their questions on moats on their tweets and speeches. 4) Finally If they don’t have an AI sidecar then it is going to be a long winding conversation. You will turn into a free tutor for them.

Published initially at https://medium.com/@mtrajan/investors-wont-fund-your-ai-startup-and-they-probably-shouldnt-d522ec9efe1e