Ask HN: Do people want to AB test LLMs?
9 comments
The challenge with A/B experiments is how you design them to have sufficient power and draw a meaningful conclusion out of them. So, you either need a big % difference between the test and the control, or you need a big number of samples. LLM apps usually don’t meet either of those two criteria. Have you ran into this with your users?
To answer the headline - No.
I find that working and adjusting my prompts and context is way higher value than A/B testing LLMs.
After all, I will never expect 100% accuracy.
I feel like they are reaching commodity status and the result quality is so similar, that it just doesn't really matter what you use.
I find that working and adjusting my prompts and context is way higher value than A/B testing LLMs.
After all, I will never expect 100% accuracy.
I feel like they are reaching commodity status and the result quality is so similar, that it just doesn't really matter what you use.
+1
There are already many prompt/LLM routers available.
We've never found value in them, for similar reasons as mentioned above.
There are already many prompt/LLM routers available.
We've never found value in them, for similar reasons as mentioned above.
Thats great feedback - thank you! I would then assume most people would end up using open source models as they tend to be cheaper with faster inference
I can save you a bit of market research and tell you that’s unfortunately not the case yet in the market today. There are a few reasons for it - the main one in my opinion being that it’s hard to measure the value vs cost of switching to an open LLM, so it’s generally perceived as same/lower value, higher cost (not in terms of inference, but in terms of overhead). What is considered however are cost saving options around the foundational models: going for the mini models, prompt caching, batch inference etc. Some tooling in that area might be interesting.
We've actually found the opposite -- Every client project has been based on GPT4 or Gemini, with one exception for a highly sensitive use case based on Llama3.1.
The main reason is that the APIs represent an excellent cost / performance / complexity tradeoff.
Every project has relied primarily on the big models because the small models just aren't as capable in a business context.
We have found that Gpt4o is very fast, when that's necessary (often it's not), and it's also very cheap (gpt4o batch is ~96% cheaper than the original GPT4). And where cost is a concern and reasoning doesn't need to be as good as possible, gpt4o mini has been excellent too.
The main reason is that the APIs represent an excellent cost / performance / complexity tradeoff.
Every project has relied primarily on the big models because the small models just aren't as capable in a business context.
We have found that Gpt4o is very fast, when that's necessary (often it's not), and it's also very cheap (gpt4o batch is ~96% cheaper than the original GPT4). And where cost is a concern and reasoning doesn't need to be as good as possible, gpt4o mini has been excellent too.
Thank you again - Great context for me
Thanks, do you think there is value in evals? / do you use evals while testing prompts?
Jumping in based on our experience in case it's helpful:
- Evals are very useful and a core part of the best practice for creating LLM apps
- There are already excellent solutions for model and prompt eval (etc.) including Parea from YC and many others
- Evals are very useful and a core part of the best practice for creating LLM apps
- There are already excellent solutions for model and prompt eval (etc.) including Parea from YC and many others
My name is Peter and I am the founder of Props. We are an AI Gateway that enables product teams to AB test models and providers using traditional business metrics (rather than evals) to measure performance.
For example: In a traditional call center, companies use NPS, CSAT, Ticket close rate ect. to measure performance. Our theory is that regardless of whether the call center is human or AI, model A or model B we should measure performance in the same way. So the customer support AI would use Props to quantify the changes they make to their AI app. They might set up an experiment that uses gpt-4o as the control and Llama 3.2 as the variant, Props will automatically split traffic between variants with our model router and allow the team to evaluate results in our dashboard.
Our ultimate goal: To build the smartest LLM router on the market
But… to build the “smartest” router, we need to solve the quality, cost and latency equation.
Cost monitoring across models is easy enough to measure, same with latency. Quality is hard because it is dependent on the business and the metrics that matter to them. So really each company needs their own custom LLM router. For that we need DATA, the way we are going about collecting the high quality labeled data is via experiments. Experimentation will provide a consistent channel for data collection that can be continually used to effectively create new fine tuned models or route to the hundreds of available models depending on the real time need.
I just soft launched this new experimentation product last week at Llama Lounge in SF (here was the demo: https://getprops.ai/demo ) and am looking for some early design partners / customers who want to work directly with us to test the product.
I also want feedback from builders. Is this useful? What would make it more useful? I am happy to set up custom demos for any use case you throw my way.
Website: https://getprops.ai Email: [email protected]
Thank you!