The "Three-Hills" Model for evaluating GPT startups

How to separate the wheat from the chaff when GPT-powered startups are popping up everywhere

TL;DR

GPT-powered startups/initiatives have to overcome 3 major obstacles to become mid- and long-term successful. This article discovers how you can analyze their use case along the dimensions of “Productivity Enhancements”, “Non zero-sum-game Value” and “Moat = Value from Context” you can classify any GPT-application into 3 different levels of potential success.

The “Three-Hills” Model

While for people who follow the tech news there is no doubt that GPT¹ technologies have started to disrupt all white collar industries at an insane pace and will have turned most jobs into assisted jobs by the end of 2024; it is currently much harder to predict who the winners of this race will be. This article aims at giving CEOs, investors and R&D budget responsibility a framework to assess the potential of GPT-based plays.

To help me as a CEO evaluate the long-term success likeness of GPT initiatives², I’ve come up with the following simple but powerful mental model:

While many applications of GPT are amazing and therefore at first sight sound like great business ideas, in my opinion only GPT-plays that surpass “Moat Mountain” are actual business opportunities in the mid- and long run.

But let’s start analyzing step-by-step:

Productivity Hill

Most initial value propositions of GPT applications today are those that increase productivity significantly. 

To be more precise: they help with tasks that fall into at least one of the following categories:

1. Speed Increase: Tasks that users could in principle perform themselves, but can now do it 10x faster in a GPT-assisted way. Examples are:

   1. Developers being able to write code fast by using the right commands and logic than looking up boiler-plate code in documentation and debugging themselves

   2. Marketing specialists, being able to write SEO optimized blog posts at 10x the speed they usually could 

   3. Students writing fast-writing essays about the french revolution instead of sitting through reading dry books and articles

   4. Looking up the best 8 things to do during your weekend trip in Lisbon, depending on your personal preferences

   5. Scientists being able to search, understand and dig into a corpus of millions of scientific papers to reach better conclusions on state-of-the art faster

2. Democratization of Abilities: Tasks that you yourself don’t have yourself the ability to perform but averagely capable people who are specialized at that not-too-niche job could. Performing them with GPT is extremely fast and free. Examples are:

   1. Writing a rhyming speech in the style of Shakespere for your best friend’s wedding that makes references to things that you have lived together.

   2. Review that long email to your boss for spelling, grammar, well-formulated arguments, tone, etc.

   3. A business owner being able to write a job ad in a descriptive and appealing way to get good applicants

   4. Consultants creating entire packages for small businesses including website coding and wording, optimized ad and marketing material texts.

   5. Formulation of grant applications with high chances of success for small research groups.

Now; before you go and start your SEO-optimiziation or Shakespearean wedding speeches startup, let’s evaluate those same ideas along the Three-Hill model.

All the above examples have cleared Productivity Hill. They all are “Level I” applications of GPT, meaning:

People will earlier or later adopt GPT technologies to solve problems on Productivity Hill and get significant value out of it. This will also significantly raise the bar for anyone offering these services professionally today.

Tug-of-War Valley

The first challenge that any Level I application of GPT will face is the Tug-of-War Valley. If the application does only create value in a pre-GPT world but will be counter-acted in a post-GPT world through opposing GPT functionalities then they are not good business ideas.

E.g. While a single marketing specialist might be able to write 100s of SEO-optimized blog articles with GPT, thousands of marketing people doing that will within a few months completely break the ability of search engines to provide useful search results based on content. Search engines (or GPT-powered search capabilities) will need to find other ways of providing useful information to the user than by indexing mediocre SEO-optimized content. Therefore it will only take a few months until SEO optimization will play tug-of-war with AI enhanced search-engine improvements.

In Tug-of-War Valley most of the benefits of using GPT will be canceled out by an opposing force also using GPT.

Value Peak

The above immediately begs the question, which companies are able to escape the Tug-of-War valley and become a Level II company. The answer is: companies that provide value outside of an existing zero-sum game: Basically any mechanism that is not meant to act as a “proof-of-work” for a human (e.g. writing essays for job applications, writing online articles, personalized emails, etc.).

Many ideas won’t make it to Value Peak:

Looking at the above examples:

• Content marketing (example 1b) won’t work in the traditional sense anymore and cease to exist in the future. Therefore SEO article writing will cease to be a useful skill very soon and tools that help SEO marketeers either have to pivot or will also quickly become a fad of the past; whether GPT-enhanced or not. 

• Classical essay-homeworks (example 1c) will in a best case scenario either disappear from curricula and be replaced with more useful teaching/learning exercises or in a worst case scenario go into an arms-race with AI-detection algorithms for teachers.

• Wedding speeches (example 2a) will soon rely even more on very personal wit and humor than on the clever text form; at least after everyone heard the 10th GPT-generated best-man’s toast.

• Your boss might not care anymore about long emails (example 2b) that propose strategy changes, since a well written long mail won’t be anymore a signal for you having spent a significant amount of time thinking it through. In fact, she might stop reading emails altogether and just listen to a GPT-formulated summary of her inbox on her way to work.

• Job ads (example 2c) and applications won’t work the same anymore. Coverletters will be the first to become useless if they are written and also screened by GPT. But even descriptions and CVs will most likely not be spared.

• However the problem of distributing grants will get solved (example 2e) in the future, if they still rely on the written word, then the tug-of-war between writing and reading software will just worsen today’s state in which agencies already started to make the process meaningless.

However, any application that generates value beyond objects made for human consumption are able to reach Value Peak:

1. Real/Usable objects: whether it is a piece of software that has been coded using GPT that makes people’s lives better, a hardware product developed using GPT enhanced requirements, etc. Each of these objects has value by themselves and are not just means to an end, independent whether they have been generated with GPT capabilities or not

2. Experiences: like a piece of music being composed for your taste and pleasure, a personalized bed-time story for your kid or an RPG adventure you can play.

3. Solving personalized problems: like providing recipes for the ingredients that you have in your fridge, helping you prepare personalized worksheets for your students, working out the logistics of your dream trip or any other problem that only you have in a certain constellation

Value Peak applications are not a fad; they will not go away anytime soon; because there is a real demand for them.

Displacement Canyon

Companies and Initiatives that assist human activities and provide clear value in a post-GPT world can reach Level II. The activities that they assist don’t just short-cut human work but create value by themselves and will therefore most likely keep existing and be valuable in the mid-run.

However, even if these applications have been thought through with specific use-cases and UX/UI, they will constantly have to compete with “generic” GPT solutions (such as ChatGPT + plugins). Why should a developer use a dedicated coding co-pilot platform instead of using a generic text assistant that can also write code?

In the past start-ups would usually get asked by investors: “what prevents the big FAANG companies from outcompeting you if they dedicated their resources to solving the same problem you are solving?”. The analogous question now will be “why should anyone use your solution and not just be assisted by a generic GPT application”?

Companies that don’t manage to climb Moat Mountain will stay in Displacement Canyon and will constantly have to compete for their users with OpenAI and similar players; most likely losing more often than winning, since the abilities of these generic assistants improves rapidly.

Moat Mountain

The only way to get to the “Basin of Success” (and therefore a Level III GPT application) is to build up a moat that generic GPT-applications (such as ChatGPT) can’t or won’t compete on directly. 

Although there might be more in the future, today it seems like there are 3 key moat-categories that startups and initiatives can implement to escape the risk of being displaced by generic GPT applications:

1. In-Context & Collaborative Features:
   Generic GPT functionalities are amazing at assisting single users with their task at hand. However, any use-case that requires collaboration of multiple people (think figma, google docs, etc.) for reviews, commenting, knowledge combination, history-tracking, dedicated access rights, etc. require a dedicated platform. When these platforms become GPT-enhanced, they become more valuable to the user, but a generic GPT assistant does not (yet?) offer context or features beyond a conversation.

2. Gated Knowledge/Data
   While any useful Large Language Model has been trained on an enormous corpus of publicly available data, gated knowledge can vastly improve the outputs of these models for specific applications. That data can be one of the following:

   1. single-client specific data: e.g. company-internal code that allows the application to suggest internally developed functions and coding styles; all legal memos written at a law firm or similar³

   2. domain specific data: e.g. databases of non-public data, such as all hardware design specifications of thousands of engineers using an engineering software

   3. complex to parse data: e.g. biomedical data or music that needs to be specifically indexed and accessed to be of value through a GPT-enhanced interface

3. Edge Computing / Offline Use Cases
   There are many reasons why some GPT use cases would need to run locally: e.g. many people might feel more comfortable that a privacy-first personal assistant that has access to all your medical data, appointments, emails, etc. did not access or store the data in the cloud. Also use-cases that enable you to interact e.g. with industrial equipment or similar, might not need to be as powerful as ChatGPT to understand verbal instructions, but are probably better kept off-grid.

Companies with a business model that relies heavily on these differentiators will very likely not only be able to attract many additional customers through the GPT enhancements, but also be able to successfully defend them.

In Conclusion

There are 3 levels of GPT-companies/initiatives. Level I applications are useful but will disappear shortly, Level II will create mid-term value but are probably not good business ideas and Level III can leverage the technology while still providing a compelling business model.

Let me know in the comments whether you agree and which (surprising?) conclusion you’ve reached when analyzing ideas with this framework.

Personal Plug:

Given my thoughts around long-term successful GPT-powered applications, it won’t come to surprise that at Valispace we just released Assisted Engineering functionalities that catapult hardware developers into this exciting near-term future today.

Since developing hardware products is a real value add and collaborative features + gated domain-specific data create moat, we are confident as a Level III company to change hardware engineering forever.
Expect JARVIS-like hardware engineering to become a reality very soon.

1. I will refer in this article to all Large Language Model (LLM) technologies as GPT. While today’s companies that are disrupting businesses through LLM almost all rely on OpenAI’s GPT or ChatGPT offer, it is becoming increasingly clear that within 2023 alternatives will be available. While this will change the named players, it does not change the underlying mechanisms and analysis in this article.

2. GPT initiatives can be startup ideas, new features or entire products, that leverage GPT functionalities at their core.

3. as of the time of writing it is not yet clear whether generic GPT applications with access to a dedicated corpus of knowledge of a single company might be sufficient to outperform dedicated platforms.