Zero-Shot Learning
00:06:18 · Adarsh Hiremath
[00:06:18] S2: Yeah. So, addressing the point about signal to noise, I would say that that is very, very use case specific. Right? So for example, in customer support, there's really only two metrics you care a lot about. Right? One is, did the AI resolve the ticket end to end? So that's ticket resolution rate. And two is, did it have a satisfaction for the customer at a rate better than a human or on par with a human? Those are two things that are very, very easy to hill climb, with the right tools, and it's a quantitative thing as opposed to a qualitative one. There are other use cases where the value judgments are just entirely, you know, vibes based. Right? And that's like a really, really important thing to get right. Yeah. For example, if you're trying to automate a consulting workflow at producing a slide deck, right, it's not, clear cut or not, like, support resolution. Mhmm. Right? So then the question becomes, how do you actually structure the preferences of the person who knows what goods looks like for that specific task in a rubric format? Right? And there, can actually break it apart into the discrete things that you would look for if you're a human evaluating the slide deck. For example, an objective thing might be whether or not the slide deck sticks to specific color formatting. Right? But a subjective thing might be whether or not the actual content on the slide deck is not verbose, easy to understand, follow some typical structure. Mhmm. Right? So the key thing is combining both of those four tasks that are unbounded and subjective, in this format.
00:07:38 · Dev Tagare
[00:07:38] S1: And are are there, like... In your journey, have you noticed some metrics that may look great on paper, but then when you go to production, when you start seeing real world use cases, they don't look as good. Perhaps one thing I've noticed over time is search as a problem space in AI land, where you get a lot of these answers really quickly, think AI mode or think whatever your favorite AI application is to get answers. Everyone personalizes to some degree here and there, but then over time, your preference is kind of distinct from mine, kind of distinct from Nancy's. Yeah. It wasn't a problem ever in SearchLand because you would just get a link back. Like, you you would get the 10 little links and it was rank rank preference across many people. But now, are you noticing that the user preference is dominating even enterprise outcomes, and gradations there?
00:08:29 · Adarsh Hiremath, Dev Tagare
[00:08:29] S2: Yeah. So, I mean, one thing that's really important is understanding the inputs into an output or or the reasoning, that actually leads to a specific, output of the model. Right? Because you could have a model that just answers correctly the first time, but it's making all the wrong decisions along the way. And then when you adapt it to a slightly different context, all of a sudden, you've got an agent that's totally broken in production. Right. So that's why the evals need to not only be over the task that it's trying to do in the output, which is the rubric that I just described, but also the trajectory that connects the two. Because you could have an LLM judge or a human evaluating whether the reasoning is sound, and if that reasoning is sound and the output is achieved properly, odds are the overall deployment is gonna go well, assuming you have, like, a reasonable distribution of all three of those things, inputs, trajectories, and outputs. To your point about whether these output metrics can be deceiving at time, the answer is 100% yes. Right? [00:09:15] S1: Yeah. [00:09:16] S2: And it's no different from metrics for like a product team, for example. Right? Where if you really, really wanted to game a metric like most product teams could, so it's about having the right right guardrails. Mhmm. Going back to the example that I, you know, pointed out earlier about customer support, you could have a high resolution rate, but the way that the agent actually communicates with the human could be totally messed up. [00:09:36] S1: Right.
00:09:36 · Dev Tagare, Adarsh Hiremath, Nancy Wang
[00:09:36] S1: Right. [00:09:36] S2: Right? It's just super ramble y, verbose, including information that... [00:09:40] S3: Or gives everyone a free flight ticket, you know? [00:09:42] S2: Or or it could give everyone a a free flight ticket, and there could be a glitch that uppays the customer out a thousand dollars for every interaction. Right? You you don't know. So you need quality control and evals over the whole chain of processes to understand whether it's actually doing a job well. [00:09:55] S1: Along the same trajectory, how do you account for, like, disaggregated learning? As an example, feedback loop in in the customer support example could be my deflection rate or resolution rate is 80%. [00:10:08] S2: Mhmm. [00:10:08] S1: But then six months later, I'm getting complaints on a specific arc or axis or trajectory, and then that is feeding back into the the overall process. Like, how does that work?
00:10:08 · Dev Tagare, Adarsh Hiremath, Nancy Wang
[00:10:08] S1: But then six months later, I'm getting complaints on a specific arc or axis or trajectory, and then that is feeding back into the the overall process. Like, how does that work? [00:10:19] S2: Yeah. So maybe the high level observation is use cases where the agent has to balance a bunch of different stakeholders' perspectives and also aggregate a bunch of different contexts are the big rock problems that we wanna solve, and they're super hard to solve because it's not as straightforward as me just prompting my GPT, claw, Gemini, whatever favorite model. Right? Because I'm always there to steer it. I have a certain set of preferences. It adds that to the memory. Right? So category a of problems with many stakeholders, a bunch of different sources of context are the really, really tricky problems to solve. Mhmm. And the way that we've we've gone about it is making sure that you can actually index the entire company's context by pulling in all the different sources. [00:11:02] S3: Mhmm. [00:11:02] S2: And that context is typically beyond a single person or a single data source. Mhmm. It's supposed to be a comprehensive and holistic view of what good of that company looks like. [00:11:09] S1: 100%.
00:11:09 · Dev Tagare, Adarsh Hiremath, Nancy Wang
[00:11:09] S1: 100%. [00:11:10] S2: And the second thing is what happens when there's actually a failure. Right? Mhmm. Typically, models can actually identify these failures if it's not confident in the answer, and then escalate it in triagic to a human. Typically, the first round of of learning updates we wanna do are in context space, not parameter space, because it's way more compute efficient. So then you can triage it to a human, update the skill. Mhmm. And if it warrants it, you could do fine tuning on model if you're using an open source model. But given how smart the models are, the former is actually usually a better approach. [00:11:38] S1: 100%. Yeah. [00:11:39] S3: So, maybe coming to, like, the technical deep dive, right, I think Dave started to unpack the layers a bit. Going back to your point around how work's actually being done, and not just sort of, you know, handcrafting prompts and sending it out there with with best hopes. Right? Is... What are some platform primitives that you're building to contextualize all of that information?