The alpha, twice a week
That was 3 min.
Get the best new moments in your inbox every Wednesday and Saturday. It's free.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
Knuckle Up with Nakul · 3 min · May 19, 2026
Decagon fine-tunes smaller models for specific customer tasks
Ashwin Sreenivas
Decagon fine-tunes smaller models for specific jobs in customer interactions, where frontier models may excel unevenly.
In this episode
Full analysis- 2:13Customer pain should shape the product before strategy4 min
- 13:09Decagon accepts coordination costs to preserve speed4 min
- 40:01Most AI companies need product-minded FDEs3 min
- 7:57Decagon brings engineers into customer sales and support5 min
- 25:52Decagon uses bar raisers to protect its hiring bar3 min
- 31:18AI writes most code, but Decagon engineers remain accountable4 min
- 35:05AI shifts engineers toward judgment and EPD leaders toward alignment5 min
- 57:19Decagon fine-tunes smaller models for specific customer tasks3 min
“you can get a model to do better at the few things that you care about in a smaller model, which is hence faster and cheaper.”
Ashwin Sreenivasat 58:15 · Knuckle Up with Nakul
Takeaways
- Frontier model performance can vary across capabilities.
- Smaller tuned models can be faster and cheaper for narrow tasks.
- Build in-house only when it improves customer experience.
About this segment
Decagon fine-tunes smaller models for specific jobs in customer interactions, where frontier models may excel unevenly. Ashwin Sreenivas says they build only when it improves customer experience; otherwise, they partner.
Ashwin Sreenivas is cofounder and CTO of Decagon.
Co-founded Decagon, the conversational AI platform valued at $4.5B, after building Helia, which Scale AI acquired.
Nakul Mandan, Host · Host, Knuckle Up with Nakul
General partner at Audacious Ventures; his show goes deep with founder CEOs on how they build.
From “The AI founder advice you won’t hear on X | Ashwin Sreenivas (Co-Founder at Decagon)”, May 19, 2026. Watch the full episode
Thanks. We'll look at and fix what's off.
Transcript
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
- 57:26
model? It's a great question. The thing that we found is that at the frontier of capability, model performance is actually pretty jagged, by which I mean, just because you have a model that advances a lot in one capability doesn't mean that it does that in every capability. And in fact, in certain cases, it might regress. Now, let me give you a specific example. For our use case, We use a whole constellation of models in each interaction with an end customer.
- 57:55
And we care about each of these models doing a few specific things. Now, I don't care about that model being good at coding. I don't care about that model being good at writing poetry. I care about a few very specific qualities. Now, if I use the frontier model for everything, in a way, I'm kind of wasting a lot of parameters and capabilities that I don't care about. And so in our experience, you can get a model to do better at the few things that you care about in a smaller model, which is hence faster and cheaper.
- 58:32
And so therefore, if we want to create the best customer experience, we should fine tune and build a lot of those models ourselves. Because we get everything better performance, cheaper and faster. I'm curious, how do you decide what to build in-house versus where to partner? So, you know, voice is clearly a core aspect of customer support. How do you decide we should build voice models in-house versus partner with existing providers?
- 58:58
Our general approach to everything here is what can give us the best customer experience the quickest, right? By which if we can get best in class from a partner, we will always do that. if we think we can do better ourselves because it is better for the end customer, then we build ourselves. I think this example of Deck Gone Labs was a great one where for the capabilities, for the kinds of models where we think, you know what?
- 59:26
A frontier model does do the best. We can't improve on it all that much if we do it in-house. Great. I would prefer to partner with someone, right? Because again, now we get to leverage their R&D spend and capabilities kind of automatically get better. Whereas in cases where I think we can build better and it is crucial to the end customer experience, those are the things that we build in-house. Like I'm not too sensitive to saying, oh, this is really important. So we must do it in-house ourselves. Everything is about where can we improve on the end customer experience.
- 59:58
In your most paranoid self, where do you and Jesse worry about, hey, if the frontier labs keep