“The AI really wants to please. The LLM as a judge wants to please. You need to tease apart every specific action the AI has taken in this response and link back to the specific instruction that gave it permission to do that. It's not just 'tell me if the AI has done anything it's not supposed to.' You have to go a significant layer deeper if you want to do this at scale.”
sputnikintelligence.com
Asked what is different about testing AI products, Greiner explains how LLM-as-judge evals for permissions and groundedness differ from ordinary deterministic tests.
…outputs across a complex customer base and trying to optimize for things like conversion, you need to be able to do those sort of like sandbox environments tests at super high scale and with a great degree of confidence. That's just not necessary when you're not directly talking with customers.
…sandbox environments. You didn't like put a new label on the skill, like, oh, you need to learn this type of online eval or this type of method gradation. What would you say are sort of those new skills that you've had to learn from the pre-AI building world to now building AI products 24-7?
I mean, I think the, I don't think the fundamentals actually change that much. I think, you know, obviously what changes is when you're dealing with a probabilistic tool, right? The type of testing that you're doing in those sandbox environments and the type of guardrails and evals that you're running change pretty drastically. You change from, hey, we have, you know, tests that our software team writes that ensure, you know, even the most advanced tests, right, ensure that when you're interacting with the product, things happen as you would expect. Now you're in a world where you need to be testing the nth possible iteration of how someone might interact with this. You're confident about what's going to come out the other end. And so like it ends up being a combination of those traditional tests that you would run, like deterministic tests that say, hey, is this output what I expected from a discretely measurable sense? And then these more advanced LM as a judge evals that are doing a greater degree of reasoning behind a much more complex. Problem like groundedness, or a much more complex problem like ensuring that the AI doesn't do anything that it…
So examining the specific traces where things go wrong is really important.
Yeah. Yeah. I think that's particularly important. And I think also, you know, examining the specific traces where things go right, right? Like I think you have the opportunity where the AI is being creative and doing things in interesting ways. And let's like the example I gave earlier, like let's…
So in our pre-call, we were talking about how you guys work. And I think this is going to be really illustrative and insightful for others because you have a unique place, right? You're not at Uber with 200 other PMs. You're a small startup. You guys get to operate and define your processes to the…