How to escape competition — Building enduring application-level value with LLMs

Sarah Tavel
3 min readApr 24, 2023

[Experimenting with the Substack/Medium cross post.]

It seems clear that over the next 5+ years, any work for which a human is a critical input into the work product is vulnerable to substitution with a software product built leveraging large language models (LLMs) and some of the related technology.

Copywriting was the first visible category of work that startups leveraging LLMs went after, with Jasper and being the first two startups going after the opportunity, and wow did they grow. There are countless other examples like this now. It’s very clear that if you are able to build a product that uses LLMs to automate a work product that has previously required hiring someone for that job, the demand will be there.

But that doesn’t mean it’s been all sunshine and rainbows for these companies. The critique that has haunted both companies (and many more) is their defensibility. This question takes two forms:

  1. If anyone with access to ChatGPT or any of OpenAI’s APIs could essentially get to the same output, you are always vulnerable to customers moving their business to whomever offers the work product at a cheaper price. As an example, both Jasper and got caught on their backfoot when OpenAI released ChatGPT, which basically democratized access to the service they were offering, spiking churn for both companies.
  2. Won’t the incumbents just add this? We’ve see Notion, Hubspot, Canva, Microsoft, etc., quickly announce GPT-driven features in their products. It becomes the classic race of either the startup figures distribution or the incumbent figures out innovation. Here, the “incumbents” aren’t sleepy companies but innovative tech companies, and the innovation necessary, while not trivial, has mostly already been done by OpenAI so it’s just a matter of fitting it cohesively into the product.

I personally think these critiques underestimate the value of focus, and don’t give the teams enough credit to execute. The reality is that very few software companies have ever had a technical moat — it’s always been about focus and execution. I do however think that the second critique speaks a bit to how we’re in a “skeuomorphic” generation of LLM-based applications where the first obvious applications use the surface level functionality of what’s possible using a new technology. So — how can Jasper/ prove the naysayers wrong? Or more generally, how do you build enduring value if you are a startup looking to leverage LLMs to create a new application?

I think this will take a few different variations that increase the odds of escaping competition:

  1. Narrowness in initial focus. One of the areas I’m most excited by right now is companies pursuing vertical application opportunities. In a world with multiple competitors seemingly focused on the big horizontal opportunities, and still rapidly evolving underlying technology, a focused competitor can win and then can expand from that position of strength. These will often involve tuning a model to a specific use case, hooking into if not replacing existing workflows (often times leveraging other ML techniques to do so), and therefore means that there will be more to the execution than a simple API call to a foundation model.
  2. Feedback loops. I’ve written about feedback loops earlier so won’t belabor the importance here. Certainly if you build an application that can leverage user engagement to improve the accuracy of your model, there will be advantages to scale and thus the ingredient to escape competition. Having a human in the loop that acts as a bit of a power user to provide feedback in the beginning is another mechanism companies use that is also effective in providing advantages to scale, provided you are able to leverage that feedback to fine-tune.
  3. Accruing data asset. I’ll confess I’m most taken by companies where a positive externality of users leveraging their LLM-driven application is the creation of a new, useful data asset that wouldn’t have been possible before at scale. In a way, this externalizes the moat outside of what’s possible with LLMs themselves and creates an even more differentiated offering that can escape competition at scale.

There is no question to me that we are at the beginning of a strong current that is going to transform the landscape of work. If you are building something here, don’t be shy. My DMs are open.



Sarah Tavel

native new yorker, SF-resident. general partner @benchmark. formerly product @Pinterest. originally blogging at