We spend $30k+ per month with OpenAI and Anthropic
Even minor prompt changes between minor model versions can have huge differences in output
We make specific tweaks to all our dozens of production prompts based on the exact model that will be used
Treating LLMs as if they are interchangeable is simply bogus
My hunch is that quite soon, LLMs will be totally interchangeable. Due to the intense competition and the fact that people are basically training on the same base data distribution.
In the tasks I'm using LLMs switching one for another makes less difference than I had predicted even.
However, I'm not spending $30k+ per month, so I guess my opinion may be less informed.
What is your use case? Could these micro-optimizations you need to do now be the result of the technology still being quite immature?
I'm working with digital twins of musicians/celebrities. But have also done more analytical stuff with LLMs.
My current side project involves working with the production company of a well-known German soap opera to help them write further episodes. The first thing we did was write a small evaluation system. Could be interesting to test with Unify.
Our prompts get very complex and we have around 3,000 lines of code that does nothing but build prompts based on user's options (using dozens of helpers in other files)
They aren't going to get more interchangeable because of the subjective nature of them
Give five humans the same task and even if that task is just a little bit complex you'll get wildly different results. And the more complex it gets the more different the results will become
It's the same with LLMs. Most of our prompt changes are more significant but in one recent case it was a simple as changing the word "should" to "must" to get similar behavior between two different models.
One of them basically ignored things we said it "should" do and never performed the thing we wanted it to whereas the other did it often, despite these being minor version differences of the same model
prompt engineering is a thing, and it's not a thing that you get on social media posts with emojis or multiple images.
is it a public-facing product?
On our roadmap, we plan to support:
- an API which returns the neural scores directly, enabling model selection and model-specific prompts to all be handled on the client side
- automatic learning of intermediate prompts for agentic multi-step systems, taking a similar view as DSPy, where all intermediate LLM calls and prompts are treated as latent variables in an optimizable end-to-end agentic system.
With these additions, the subtleties of the model + prompt relationship would be better respected.
I also believe that LLMs will become more robust to prompt subtleties over time. Also, some tasks are less sensitive to these minor subtleties you refer to.
For example, if you have a sales call agent, you might want to optimize UX for easy dialgoue prompts (so the person on the other end isn't left waiting), and take longer thinking about harder prompts requiring the full context of the call.
This is just an example, but my point is that not all LLM applications are the same. Some might be super sensitive to prompt subtleties, others might not be.
Thoughts?
It's already hard enough to get consistent behavior with a fixed model
If we need to save money we will switch to a cheaper model and adapt our prompts for that
If we are going more for quality we'll use and more expensive model and adapt our prompts for that
I fail to see any use case where I would want a third party choosing which model we are using at run time...
We are adding a new model this week and I've spent dozens of hours personally evaluating output and making tweaks to make it feasible
Making it sound like models are interchangeable is harmful
I would also not try to emulate DSPy, which is a massively overrated bit of kit and of little use in a production pipeline.
For something like a support agent why couldn't the company just choose a model like GPT-4o and stick with one? Would they really trust some responses going to 3.5 (or similar)?
Still, I agree that routing in isolation is not thaaat useful in many LLM domains. I think the usefulness will increase when applying to multi-step agentic systems, and when combining with other optimizations such as end-to-end learning of the intermediate prompts (DSPy etc.)
Thanks again for diving deep, super helpful!