A Playground for LLM Apps: How AI Engineers Use Humanloop
In the evolving LLM app stack, a British company called Humanloop has — perhaps accidentally — defined a new category of product: an LLM “playground.” It’s a platform where developers can test various LLM prompts, and then deploy the best ones into an application with full DevOps-style measurement and monitoring.
To understand exactly what Humanloop offers developers, and how it became one of the leading “playground” toolsets, I spoke to Humanloop co-founder and CEO, Raza Habib.
I first learned of the term “playground” in the LLM app stack diagram created by Andreessen-Horowitz (a16z). But what does it mean and where did the term originate?
Via a16z; Click image to view full-size.
Habib, who holds a Ph.D. in machine learning from University College London, explains that it derives from OpenAI.
“When OpenAI first released GPT-3, they just wanted to have an environment where people could go try the model — and they called it the playground. And so […] the name has stuck around. But I think the point is that it’s an environment to interactively try things with different models.”
Habib also noted that a16z didn’t initially know where to place Humanloop in its stack.
“I think we could have belonged in a couple of different places on that diagram,” he said. “But at its core, we help developers evaluate and take steps to improve their prompts and AI applications.”
Let’s take a step back. As Habib explained it, LLM applications start with a base model — such as GPT-4 or Claude — or maybe your own large language model. To begin creating an application you need a “prompt template,” which Habib described as “a set of instructions to the model, with maybe gaps for input.” You then “chain together” all of this with other models or with information retrieval systems to build out a whole application.
Where Habib and his co-founders spotted an opportunity in this process was in collaboration — helping technical users work with non-technical users to try different prompts and evaluate them.
“What we’ve found, speaking to people working on this early on, is [that] it’s very iterative,” he said, regarding the process of building an LLM application. “It requires collaboration between technical and nontechnical people to find good problems and get these systems working well. And evaluation is really hard, because it’s very subjective.”
Diagram via Humanloop
Use Cases and How It Works
One of Humanloop’s customers is Duolingo, a popular language education application. As with many other tech companies over the past year or two, Duolingo has been busy adding AI to its core product. A recent blog post explained that it uses AI in a variety of ways, including helping its staff create lessons and “build courses faster.” Writing prompts are at the core of this:
“Here’s how our AI system works: We write a “prompt,” or a set of detailed commands, that “explains” to the AI model how to write a given Duolingo exercise.”
Duolingo is careful to emphasize that the ultimate responsibility for its lessons and courses falls on its human instructors. Nevertheless, it’s clear that AI is helping a lot — both with template design and to “fill in the blanks.”
Where Humanloop comes in is to help Duolingo get the right type of content out of the LLMs.
“It [the content] obviously needs to be appropriate to the learner, the right tone, the right language, vocabulary that’s appropriate, etc,” Habib explained. “So, it’s not trivial to take the base models and actually get them to do what you want. And so what we provide is a set of tools for iterating on, collaboratively, your prompts and your workflows; measuring performance in production; and then also being able to monitor and evaluate things over time.”
Typically Humanloop is used at the prototyping stage. A team of people will open Humanloop (which is browser-based) and they will see “a playground-type environment.”
“They can try out different models, they can try out different prompts,” Habib continued. “They can include that in a sequence or workflow. They work on that till they get to a point where it seems to be working reasonably well [and] now it’s time to go in and try it out more seriously, beyond just eyeballing things. They’ll then typically run more quantitative evaluation, and so we have the ability to set up evaluation functions. They’ll deploy that to production, and they’ll monitor how well it’s working — including being able to gather end user feedback.”
A similar workflow happens when doing tweaks or testing out new prompts, so it’s an iterative process that doesn’t stop after the application has been deployed.
Playing with Others
I asked whether Humanloop can be used in tandem with other products in the LLM app stack, such as the orchestration framework LangChain and vector databases like Pinecone?
“It integrates natively with LangChain [and] a couple of others,” he confirmed. “So you can switch on an environment variable in LangChain, and then you’ll automatically start getting logging and monitoring of your applications in Humanloop. So it’s really like a one-line code change and then suddenly you can see what data is flowing through, and start gathering feedback and take actions to improve and debug.”
Habib noted that Humanloop has a feature similar to OpenAI’s functions, which it calls “Tools.” This allows users to “connect an LLM to any API to give it extra capabilities and access to private data” — for example, to connect to a vector database. But Habib cautioned that Humanloop isn’t an orchestration framework like LangChain.
“We believe that that’s best done in code,” he said, regarding orchestration. “We’re primarily there to manage the prompt engineering and then evaluate and take steps to improve those models.”
Advice for AI Engineers
The primary users of Humanloop are developers. With the current popularity of LLM applications, I asked Habib what advice he’d give to developers who want to do more work in this area.
“In terms of new skills you want to learn, I think having an awareness for how the models work and an appreciation that this is now stochastic. So if you haven’t had any experience with machine learning before, and you’re coming into it, you’re probably coming from a world in which software is deterministic — [where] you can write unit tests and it always does exactly the same thing.”
With LLMs, though, software isn’t necessarily deterministic. So learning to deal with that randomness and developing an intuition about the limits of LLMs is important, in Habib’s view. Which, of course, is where an LLM playground comes into play.
