Zero-Shot Learning
00:17:33 · Dev Tagare, Adarsh Hiremath
[00:17:33] S1: And and in this, again, we we come back to the question of, like, time is not continuous. As an example, I may pick up a task today, I may do it over the next five days, abandon it randomly, pick it up in two weeks again, so on and so forth. And in your process of discovery, maybe you find this as a high impact task to have actually done. So help us understand how do you think about that process of discovery and then tying it back to disaggregated steps, if you will. That models to your point earlier, basically an agent trajectory. We're moving from agent is solving for a goal to agent is sticking to a trajectory to get to the goal side of paradigm. How do you think about that? [00:18:19] S2: Yeah. So maybe breaking apart that specific thing of sometimes humans will do a task in one sitting, they might wait five days. There's a couple of reasons for why a person might do that. The one is that the actual task is not that important. Right? And you need to reprioritize and do something different, or you just don't have the bandwidth. Or or or the second thing is that you just actually didn't get around to it, even though it is important, which actually is also a bandwidth problem. The beautiful part of agents is that they don't have these constraints. Yeah. Right? They can prioritize. They can operate at exceptionally high throughput. So when there are a bunch of different tasks that can be taken [00:18:55] S1: Right.
00:18:55 · Dev Tagare, Adarsh Hiremath
[00:18:55] S1: Right. [00:18:55] S2: The agent should be able to do them all simultaneously, assuming they're high value tasks that would that would be important in an enterprise context. The piece that I think is unique and has not been solved is the continuous learning and updating. Right? How do you make sure that the agent doesn't have, like, a long feedback loop before it understands that it made a mistake or even understands that it could be doing better. Right? Yep. That continuous learning loop involves a bunch of different things. It involves triaging tickets to humans who know what good looks like, if that's appropriate. In some cases, it involves automatically understanding its failure from LLM as a judge. In some cases, you could actually even look at the telemetry of the API call, and then you can have that automatically improve the agent. So that continuous learning as a problem space, I find, is is very, very interesting and unsolved.
00:19:39 · Nancy Wang
[00:19:39] S3: You know, when I would say humans are doing workflows, right? You pause because you might not have actually all the context, even of Exactly. The Or like, hey, you realize I need to tag in so and so because they own this particular work stream. For the agent itself, right, because oftentimes we see, you know, hallucination. For example, internally, we have a no code agent builder that allows our GTM teams to pull in, like, context from across Drive, right, Slack, Notion, so on and so forth, for their presentations. For fields, oftentimes, that we find that the agent doesn't know, it very confidently just makes up data. Right? So, do you prevent that, where a natural human might pause, but the agent might just continue and finish that task?
00:20:21 · Adarsh Hiremath
[00:20:21] S2: I mean, the key thing is just citing the relevant company context. Right? If the agent is saying a fact and it is unable to cite the company context, it should automatically stop. Right? That's a functionality we already have in the Right. Merkur platform. That's just like maybe level one. Level two is if it's able to cite the context but gets the wrong interpretation, you need to either have a human review the trajectory to create the eval or have an LLM review the trajectory. Because some LLMs might be, unable to get the right answer, right, or even able to get the right answer. But they might even be better at reviewing the trajectory. Right? So it's like you could arrive at the wrong answer, but it could know that it arrived at the wrong answer and why, and it could also arrive at the right answer and know that the reasoning was wrong and why. So you typically wanna have multiple levels of review on the actual trajectory of the agent.
00:21:06 · Dev Tagare
[00:21:06] S1: Yeah. Super cool. And in your observation, lot of these incorrect answers or hallucinations usually tend to happen in what you would describe as conversational agents. I'm asking a question, I'm getting an answer back quickly in the pursuit of quick answers, often grounding is ignored sometimes, citations are ignored sometimes, the models underthinking, overthinking, what have you. Are you building towards a world where quality or precision takes over conversational kind of experiences? And there's a curve there. Right? Like, you trade latency for accuracy, so to say, but then eventually you wanna bring accuracy to latency.
00:21:51 · Adarsh Hiremath, Nancy Wang
[00:21:51] S2: Right. Uh-huh. Yeah. So so addressing that, that's one of the reasons I'm most interested in in evals as the foundational infrastructure investment that makes all of this possible, which is it allows you to select the best model for a specific use case. There is no quantitative way or data driven way to make that selection unless you have an eval suite that is representative of the job to be done and is comprehensive across the task and the trajectory and the output. That's why we're so so so excited about evals and what they mean for the future, because the models are getting to the point where the real bottleneck is applying evals throughout all the economically valuable parts of the economy and then... And steering the models in in the right way. To your point about conversational interfaces, versus maybe using a agent as the backbone of, like, a broader system. I would say that in conversational interfaces where the person is working directly with the agent Mhmm. Mistakes are are actually a little bit more pronounced and easy to to to see. Right? It's like the agent said something that doesn't quite make sense or just weird. Right? There... There's that. But I actually think the more dangerous failures are the ones happening behind, the scenes of these broader systems where you don't have a conversational interface to diagnose the failure. And often these failures happen silently, and often they have much, much larger consequences than one would anticipate or want. [00:23:07] S3: Right.