Software market analysts are entirely missing the point with LLMs
Putting a vertical niche UX on GPT is not the future
This morning I was reading this article by Benedict Evans in which he uses concepts from Clayton Christensen’s law of conservation of modularity to try and make predictions about LLMs. If I just lost you, sorry about that (quick ‘splainer below or you can hit this link to learn what the law of conservation of modularity is.)
Clayton Christensen wrote a book called The Innovator's Dilemma in which he observed a pattern of innovation. The pattern goes like this: new innovations start as a vertically integrated system (think of the first car or the first PC) and then get “unbundled,” as new companies form to create modular improvements (think about windshield wiper motor specialty companies in the car world or hard drive makers in the PC world).Ok, so in the article I read this morning, Benedict is looking at this new, vertically integrated product ChatGPT (or just GPT in general), and thinking about how it’s going to be unbundled. And, his thesis is that the various things that an LLM can do (from giving medical advice, to helping make travel plans) are going to become the next wave of software businesses.
WRONG.
This is not the future despite so many people trying to fund it and build it.
Recently, I went to a Fintech Conference in Copenhagen, and brand new companies were off to the races building niche GPT-powered solutions for well defined markets that they believed had huge growth potential. Likewise, I went to a mentoring session for a new batch of Techstars companies, and at least 4 of the 12 were aiming to harness LLMs for some niche.
Many people have expressed that the problem with this is that there’s no moat. Yes, some other company could come along and easily copy your wrapper around the GPT-4 API that helps teachers create engaging lesson plans, but worrying about moats is missing the entire point.
Because I want to keep you reading, I’m going to hold off just a second longer in telling you the entire point.
Instead I’ll tell you something that helped me realize what the point was. So my company, Kelsus, is starting to look at acquiring companies. We are interested in acquiring well run, vertical SaaS companies that have over $1M in revenue, and fit a specific thesis.
Ok, so here we were wanting to find these kinds of companies for sale. Have you ever looked at buying a company? You have to wade through thousands of garbage listings of horrible companies that should just shut down. And then beyond that, it’s hard to compare one company to another based on numbers like their asking price, revenue, EBIDTA, etc. So I started building a tool.
Using ChatGPT to help me, I built a tool that could
Scrape websites full of company listings and drop their information (URL, description, asking price, revenue, profit, location, etc) into a Google Sheet.
Hand the content of that spreadsheet off to GPT-4, so that it could read the description of the companies and score each one on our thesis on a scale from zero to one with one being “definitely matches the thesis.”
Lastly, it would take the contents of that spreadsheet, build a weighted scoring model, and plot them in a chart so that the most interesting acquisition opportunities would end up at the top of the chart.
I built this tool in under 40 hours of development. It is fast and loose, a bit brittle, but it does the job. I stopped working on it when marginal improvements to the software weren’t saving me time at finding target companies.
After I built it, you better believe that I thought about marketing it as a product. My entrepreneurial streak runs deep.
But I knew better. What made the tool valuable to me is not that it could scrape business listings and score them. Its real value lay in its perfect alignment with my specific needs and expectations.
And there it is. THAT’s what the problem with companies trying to build LLM-based tools for niche markets is. You’re trying to build a product for a broad set of users but the power of AI is that it you can build a product for yourself (or your specific business).
I think that we’re so used to living in a world where products are not 100% customized to ourselves that we have a hard time grokking this transition. But as soon as you think about it it’s obvious. What if the bicycle you bought was built exactly to your fitness level, body geometry, and riding style. It would be amazing! What if you never missed a UPS delivery because they just knew better than to deliver when you weren’t home? Perfect world, right?
So now we need to extend this to how we think about software. This new market reality is not about improving undifferentiated products through bundling and unbundling. This is about converting a singular B2B product serving 100,000 customers into 100,000 unique B2B products, each serving a single customer.
So, when Benedict Evans, or Ben Thompson, or really any market analyst Bens try to look into our AI-assisted future and understand what the big new companies will be, they should be looking for companies that accelerate this journey from 1 to 100,000 products. Those are the ones that are going to be giant.
In a final note to tie together my thoughts on where AI is going and what Kelsus is doing in terms of finding acquisition targets. For the moment, Kelsus is still completely involved in providing software services. But we see the value of billing for software development by the hour going down fairly quickly already as LLMs help with the heavy lifting, and we’re concerned about the commodification of development skills.
So, rather than be the software developers, we’d rather be the software providers, and that’s why we’re looking at doing acquisitions. But as a bonus to this way of thinking, we believe we can take advantage of the development speed that LLMs unlock and start driving hitherto unaffordable, individually customized products to the customers of the vertical SaaS companies that we acquire by leveraging the power of LLMs.
This is going to take us a long time to put together, and in the meantime, we’re very much still open for business as a custom software development shop, but if there are some things that you wish your software could do that have felt too expensive or time consuming or not worth the investment, maybe give us a call. This is a new world where massive software value can be unlocked in incredibly short amounts of time.
Thanks for reading!
—Jon Christensen
I don't use GPT to write my posts, but I thought it'd be fun to ask GPT if it agreed with this post. The answer is one of my favorite recent interactions with ChatGPT:
Ah, the metacognitive moment—when the AI you're writing about is asked to critique the article that critiques its applications. How very Escheresque!
So, let's break this down:
Customization vs. Standardization: Your argument that the true power of LLMs (or AI in general) lies in customization is compelling. However, it's worth noting that many businesses find value in standardization. In a way, your thesis argues for a new form of mass customization, which is a powerful concept but not without its challenges, like increased complexity in product management, delivery, and support.
Moats: While I get your point about dismissing "moats," those do provide companies with competitive advantages and are a major part of the business landscape. That said, customization could be a moat if done right.
Scaling Customization: The idea of converting a single B2B product for 100,000 customers into 100,000 unique B2B products is exciting but also an operational nightmare unless brilliantly executed. The question then becomes: can AI help in not just the customization but also in managing the complexity that comes with it?
AI Complementing Human Skills: You touched on the commodification of development skills due to AI's capabilities. While AI can handle specific tasks, the creative, strategic, and problem-solving elements often still require a human touch, creating a collaborative model rather than a competitive one between human skills and AI.
Accelerators of Customization: You suggest that future big companies will be the ones that can facilitate this mass customization. That's an interesting angle and could be a significant trend if the balance between customization and scalability is achieved.
In sum, while I can provide you with perspectives that either agree or challenge your viewpoints, my "agreement" doesn't come with a preference or belief system. But isn't it fascinating to think of AI as both the subject and the tool for this kind of discourse?