Revenge of the GPT Wrappers: Defensibility in a world of commoditized AI models
andrewchen.substack.com
andrewchen.substack.com
A "wrapper" will always be better than the the foundation models so long as it can do the domain-specific pre-generation ETL and data aggregation better; that is the true moat for any startup delivery solutions using AI.
Your moat as a startup is really how good your domain-specific ETL is (ease of use and integration, comprehensiveness, speed, etc.)
The ones derided were those claiming to be 'open-source XY' while being a standard tailwind template over an OpenAI call or those claiming revolutionay XY while 90% being the proprietary model underneath. I am not sure how many were truly innovative that weren't cloneable in a very short time. Using models to empower your app is great, having the model be all of your app while you pitch it otherwise is to be derided.
I asked them, "Well, it's open source. Instead of making a bunch of adapters, couldn't you just copy the code you want?"
Turns out the whole agent was 11 files or so. The files were about 200 lines. Over half were just different personas to do the same thing. They just needed to copy one of the prompts and have a mechanism to break the loop.
The funny part with open source is nobody reads the code even though it's literally open. The AI pundits don't read what they criticize. The grifters just chain it forward. It's left-pad all over again.
Both essays make convincing points, I guess we'll have to see. I like the Uber analogy here, maybe the winners will be some who use the tech in innovative ways that only leverage the underlying tech.
e.g. OAI Operator, Anthropic Computer Use, and Google NotebookLM.
There’s plenty of applications to build that won’t easily get disrupted by big AI. But it’s important to think about what they are, rather than chase after duplication of the shiny objects the big companies are showing off.
https://interjectedfuture.com/the-moats-are-in-the-gpt-wrapp...
Yes. It will become a duopoly, where the leading frontier model holds >90% market share, and most useful products will be built around it. With the remaining 10% being made up by large portions of the other big vendors, and then everyone else for niche cases.
The idea of picking and choosing between individual models for each specific use case is going away rapidly as the top ones pull away from the pack, and inference prices are falling exponentially.
It's actually not that easy to copy/paste AI agents, prompts take quite a lot of tweaking and it's a rather slow and manual process because it's not that easy to verify that they're working for all the possible inputs. This gets even more complicated when you get a number of agents in the same application and they need to interact with each other.
Besides that, you quote "imagine it becomes...", it is a fair to assume that these technologies will become better.
It's capable of creating CRUD apps from scratch more or less by itself, and I can see how in this area we soon might get to a point where you can get your own clone of a lot of apps up and running in 30 minutes.
But I imagine a lot of future value we might see created will come from:
1) specialized prompts - looks simple but I don't think it is, especially if you have 100s of them in your application and you have complex logic on how they interact between each other, you're using different models for different parts of your application based on their strengths, etc
2) access to structured data you can connect your agents to
3) network effects - app that is mostly used gets better just by using the usage data (the article did talk about network effects)
I don't think it's really easy to replicate these 3 factors. The article is also mentioning some of this, I'm not really arguing with that, just pointing out that I don't think it will be that simple to c/p full applications.
It’s not “trying” that matters. What matters is testing. But nobody is testing LLMs… Or what they call testing is mostly shrugging and smiling and running dubious benchmarks.
Both Apple and Google are doing a poor job of integrating AI capabilities into their Operation Systems today. Maybe there is room for a new player to make a real AI-first Operation system.
An AI-first pane of glass (OS, browser, phone, etc.) with an agent that acts in my behalf to nuke ads, rage bait, click bait, rude people on the internet, spam, sales calls and emails, marketing materials, commercials, and more.
If you want to market to me, you need to pay me directly. If you want to waste my time, goodbye.
OS is generally expanded to Operating System, not Operation System, in English
(Not saying it's a bad or good thing, nor saying AI is comparable)
- Ray Kroc – Turned McDonald's into a global fast-food empire.
- Howard Schultz – Scaled Starbucks into an international giant.
- Michele Ferrero – Created Nutella, Kinder, and Ferrero Rocher, making his family billionaires.
Water:
- François-Henri Pinault – Controlled Evian via Danone.
- Antoine Riboud – Expanded Danone into a bottled water empire (Evian, Volvic).
- Peter Brabeck-Letmathe – Former Nestlé CEO; Nestlé owns Perrier, Pure Life, Poland Spring, etc.
Electricity:
- Warren Buffett – Berkshire Hathaway Energy owns multiple utilities.
- Li Ka-shing – Built major energy holdings through CK Infrastructure.
- David Tepper – Invested heavily in power utilities via Appaloosa Management.
Enron did!
[1] https://www.forbes.com/profile/sarath-ratanavadi/?list=rtb/
Not everyone will have the training data that demonstrates precisely the behavior that your customers want. As you grow, you'll generate more training data. Others can clone your product immediately... but the clone just won't work as well. In your internal evals, you'll see why. It misses a lot of stuff. But they won't understand, because their evals don't cover this case.
(This is quite similar to why Bing had trouble surpassing Google in search quality. Bing had great engineers, but they never had the same data, because they never had the same userbase.)
Do you think it's less work than just making GPT B and why? What quality in the system (inductance aside) is this simply additive?
My strawman reads as "wishing for fairytales" basically. But this strawman to me, is the reductive intent inside the article. "Ask a GPT to perform like another later different GPT" epitomises magical thinking.
Why bother training if the recursive application is that simple? Because... it's not that simple.
Imagine a new UI/UX for a CRM. Completely redesigned from the ground up.
Multiple GPT wrappers in a single product. All wrappers working together to achieve a single goal.
Also throw in agents.
And distribution matters, if some app goes viral, it has a high chance to succeed and beat the current incumbent.
I'd be more worried about platform dependency at this point. The ol' adage about building your business on someone else's API applies, doubly so in the current geopolitical climate. All those hot AI startups are but a single executive order away from losing half their market, or getting erased from existence altogether.