OpenAI will “steamroll” startups; what should we build as the models get better?

Create Resumes at CVGist
3 min readMay 22, 2024

Wouldn’t it be nice if AI providers only wanted other companies to simply use their models and develop niche apps into business domains the providers will never play in. Way smarter than trying to be everywhere on the market, right? After all it’s the common platform strategy.

Realistically, AI providers will steamroll everyone in their way & on their quest for holistic market adoption. There are repeat patterns in technology platform & ecosystem evolution that we’re in the early phases of; for example AWS’s rise to platform dominance.

Today though is a bit different as intelligent and agent automation is a complex variable that now breaches into business and niche domains (something AWS didn’t do as an infrastructure provider). In other words, how do you build something when the big players are building something else aimed to generally solve everything? A generalized agent is intended to do just that after all.

What do we build then?

This is a tough question to answer today and predict the answer to for tomorrow. We see new forms of advanced capability every quarter. Just last week OpenAI was demoing a live conversation with AI, exemplifying a multitude of capabilities such as real time language translation; Duolingo’s stock dropped the following day. Then, the day after, OpenAI demoed automated Google Sheets analysis and creation (overshadowing at least 5 developer-focused “data analysis agent” startups… and evening signaling a potential career-reduction scenario for data analysts themselves).

In my brief opinion, you should build to enhance complex workflows that LLMs can’t yet directly interface with. Where you’re not only orchestrating advanced LLM usage, but also integrating into business processes and other systems. The more glue you’re building, the safer your startup is as AI advances.

Example: Resume Generation with AI

A simple example: Resume Generation… literally everyone can go to ChatGPT today to make resume text. The market is also flooded with tons of resume generators. But what is missing is using AI to generate not only the full text, but also the fully formatted Word document — not a json populated web ui or pdf — and in a way where users don’t to have to fill out forms. Powerful AI Interface + Limited User Involvement + Integrations is the key imo. CVGist.com is a simple example of this for resumes. All a user has to do is type in a simple gist, and CVGist uses ChatGPT + a bunch of document integrations to generate actual words docs (90+ resume templates formatted). The simple glue we built here is generating word documents… which wasn’t entirely simple. The word doc generations is what we’re expanding on business idea wise; beyond resumes, which even still we must be adept to Microsoft’s own ability in innovating word processing with AI.

Of course, even this isn’t a great one in the long run. We’ll soon have agents creating highly structured documents for us across various styles. The point is to seek to build the integrations.

More glue is the key

More advanced integrations for the resume example, or the “glue”, might include applying to multiple jobs for the user (involving web apis, perhaps headless browser automation, etc). But OS and Browser providers will integrate agents natively, and we shouldn’t be surprised if OpenAI releases their own agent browser one day. In such a scenario the browser based agents/assistants would be powerful enough to conduct any basic web task (such as filling in forms) for the user in a native browser conversational experience (I dont think it’s a good idea for startups to try and build here).

Therefore, ensuring your integrations are future-proof is an additional step in idea validation. Find and work on systems and business workflow process that are hard to breach from a simple integration perspective. These should be complex workflows beyond simple UI/API interactions (because again the providers are headed towards generalized platform agents).

So what’s a idea that is future proof and won’t get steamrolled by Altman? Any answer here is worth a lot of dough.

--

--