I have a slightly different take on this

May 9, 2025

I have a slightly different take on this. (Or maybe it's not a different take, just a narrower view.)

The underlying models and infrastructure you use are not going to differentiate your product.

But things are changing so fast all the time right now, and that's not slowing down: better models, evolving best practices, new techniques for reliably implementing complicated workflows. We're still only at the "the agents are pretty good" stage. There's a lot of headroom.

My guess is that we have at least another couple of years ahead of us where *speed of iterative improvement* will determine the winners.

Did Gemini just leapfrog GPT-4o in a way that matters substantively for your use case? How fast can you leverage that? How fast can you switch back when OpenAI counters?

Did Llama just get good enough that you can offer your enterprise customers a fine-tune that is 10x cheaper *and* lets them keep their data inside their own VPC? How fast can you deliver that and demonstrate that the performance is the same on task-specific evals?

Can you offer your enterprise customers the same performance in Spanish, French, Hindi, and Tagalog as in English? (This is not easy, today.)

I talk to a lot of developers every day, both @trydaily customers and members of the broader @pipecat_ai community. Going from "pretty good" to "really good" is a jagged frontier and a moving target. But doing it unlocks much bigger opportunities and being able to do it repeatably feels to me like the closest thing there is to a moat in this space, right now.

Hadley Harris@Hadley

Probably the hottest thing raising seed rounds this year: voice agents — including some in our own portfolio.

Startups are showing you can grow very fast by automating calls, replacing labor, and capturing unstructured data at scale.

But there’s a catch… 🧵
1. While these