At Relume, we're currently not planning to fine-tune our own LLMs, and we believe this is the right strategy Here's why: 🎯 Our focus remains on delivering new features to our customers. 🚀 We can quickly adopt new models as they become available. 💡 It simplifies our deployments, saving us time and resources. We suspect this will change in the future as we scale and as the landscape matures. What's your take? Is it better to train your own models or use APIs?
I'd guess it depends on the problem you are solving. models are also becoming easier to train, build for scale from the start, rework is a pita.
I’m guessing that companies that only work on models will be able to provide a much better solution, then you can tailor it for your needs
For where Relume is at right now Daniel James Slater it definitely makes sense to just focus on moving quickly and worry about fine-tuning later. I suspect though in about ~12-months as you continue to grow, the need for a more performant LLM will increase so you'll find yourself starting to fine-tune them for specific use-cases across your product.
If it's not your core business, I definitely wouldn't care about this now. Focus on delivering new features to your customer and you can worry about it later.
Staying focused on delivering value to customers is key! 👍 Your strategy sounds solid.
Great strategy! It's wise to focus on delivering new features to customers and quickly adopt new models as they become available. I think leveraging APIs offers a more efficient pathway for numerous companies. The landscape is evolving at such a rapid pace that models from just last week are becoming obsolete.
Yep that's been my take too. Better to work on frameworks and architecture. The space is moving so quick you need to understand the likely future state that prod will launch into. So being model agnostic is 100% the right attitude.
founder @ start2scale ⤴️ | Global Talent Acquisition for hyper-growth 🚀
10moNice. Spoken to many startups that: start with a quick to market wrapper, early adopters + PMF, interview, interview, interview, analyse opportunity cost of custom LLM, then probably continue with wrapper and deploy capital to hack growth. Pivot somewhere if wrong. Can’t hate on it.