10 Quotes on AI Agents From Harrison Chase, Co-Founder and CEO of LangChain

Harrison Chase is the co-founder and CEO of LangChain, an open-source framework that enables developers to easily build applications powered by large language models (LLMs). Chase launched LangChain in October 2022 while working at the machine learning startup Robust Intelligence. The project quickly gained popularity among developers for its modular abstractions and extensive integrations that simplify the process of creating LLM-driven applications.

Prior to founding LangChain, Chase led the ML team at Robust Intelligence and the entity linking team at Kensho, a fintech startup. He studied statistics and computer science at Harvard University. As CEO, Chase has overseen LangChain’s rapid growth, with the company raising over $30 million in funding at a $200M+ valuation within just a few months of incorporation in 2023

Given his important contributions to the field of AI, and in particular, AI agents, let’s take a look at 10 quotes from Harrison Chase on the topic:

1. “I don’t think we’ve kind of nailed the the right way to interact with these agent applications. I think a human in the loop is kind of still necessary because they’re not super reliable. But if it’s in the loop too much, then it’s not actually doing that much useful thing. So, there’s kind of like a weird balance there.”

In this excerpt from a presentation with Sequoia Capital, Chase highlights the challenges of designing effective user interactions with AI agents. He emphasizes the delicate equilibrium needed between human oversight and agent autonomy to ensure reliability while maximizing the agent’s utility.

2. “Agents are like digital labor – capable of automatically browsing the web, navigating our files using our applications, and potentially even controlling our devices for us.”

During his TED talk, Chase introduces the concept of AI agents as digital entities that can perform tasksautonomously, such as web browsing, file navigation, and device control. He likens them to a form of digital labor.

3. “We’re basically constantly using a variety of different tools to help us with a given task. This is where agents are a bit different – instead of us using those tools, we just describe to an AI what the task is and what the end goal is, and then then it plans which tools it needs to use and how to use them and then it actually does it on its own.”

Chase draws a distinction between the traditional approach of humans using tools to complete tasks and the AI agent approach. With agents, users simply describe the task and end goal, and the agent autonomously selects and uses the necessary tools.

4. “Not only can they complete the task much quicker than we can, but in theory, we wouldn’t even need to know how to use these tools in the first place.”

Expanding on the benefits of AI agents, Chase notes their potential to complete tasks faster than humans. He also suggests that agents could eliminate the need for users to have prior knowledge of the tools required for the task.

5. “I think there’s probably like two places where it’s going. One is like more generic tool usage, so having, you know, humans specify a set of tools and then having agents use those tools in kind of like more open-ended ways.”

In an interview, Chase discusses future directions for AI agents. He envisions agents using user-specified tools in more flexible and open-ended ways as one area of development.

6. “I think the idea of like long-term memory is really interesting so having agents remember things over time and kind of like build up knowledge.”

Chase identifies long-term memory as another key area for AI agent advancement. He’s intrigued by the potential for agents to accumulate knowledge over time and leverage it to inform their actions and decisions.

7. “We kind of like condensed that into information and so I think that’s a really interesting step in this idea of like more personalized agents that know more about you.”

Elaborating on the concept of personalized agents, Chase explores how agents could condense information from a user’s interactions and preferences over time. This would enable a more tailored and individualized agent experience.

8. “I think there’s a big pain point that this is solving which is like for all these generative models, it’s really hard to evaluate them.”

Chase discusses the challenge of evaluating generative models. He suggests that AI agents could potentially help address this pain point.

9. “And that’s because you aren’t producing like a single number that you can do like MSE on or accuracy or something like that, you’ve now got like these, I mean at the very least you’ve got like a natural language response.”

Chase elaborates on the difficulty of evaluating generative models, noting that their outputs are often natural language responses rather than easily quantifiable metrics like mean squared error or accuracy.

10. “So I think that’s a that’s an area that yeah, I mean, we’re both extremely excited about, I think, is using language models themselves to evaluate language model outputs.”

Chase expresses enthusiasm for the idea of using language models to evaluate the outputs of other language models, seeing it as a promising approach to addressing the challenges of generative model evaluation.