AI capabilities seem to grow faster than people can build applications on top of them. And often the companies best places to leverage AI are the companies building the models themselves.
For example: something like Operator offered directly within ChatGPT makes 100s of startups building ‘LLM + RPA’ obsolete. Claude Code may in time compete directly with Cursor and the likes.
As product builders, how should we cope with this? I think we should start positioning ourselves so that better, stronger AI makes our products more useful, not less! We should plan for increasing capabilities, predict the future and build for it; instead of reacting and trying to solve yesterday’s problem.
Here’s a concrete example. When we were building our WhatsApp filing flow for the first time, we had a very ‘guided’ flow where our product would run the chat in a pre-defined sequence. This didn’t work very well — it couldn’t handle unexpected responses from the user, it was very fragile. The older models (GPT 3.5 generation) were not powerful enough yet.
We still took a bet that AI will get more powerful and cheaper. We understood that with a more powerful model, the AI could take better control of the conversation — it could drive the chat towards the right outcome without us needing to define each step in painstaking detail.
This bet paid off. Models continued to improve (GPT-4, 4o, 4o-mini came out). They were better able to react to unexpected messages, and guide the user towards a successful filing.
Initially, the new models were very costly. But models continued to improve, and we found some really creative ways to optimise costs. We were able to reduce costs 600x while maintaining performance.
My takeaways here are:
- Bet on AI getting better (and cheaper) over time.
- Don't constraint yourself to current capabilities. Predict the future and actively plan for capabilities getting stronger.
- Think deeply about how our products can do much more for the users in an age of ubiquitous AI.