LANGCHAIN — Can OpenAIs Bet on Cognitive Architecture Succeed?
Software is like entropy: It is difficult to grasp, weighs nothing, and obeys the Second Law of Thermodynamics; i.e., it always increases. — Norman Augustine
OpenAI’s recent developments, including the release of the Assistants API and GPTs, signify a significant bet on a particular cognitive architecture. This architecture focuses on agent-like systems powered by large language models (LLMs), aiming to drive applications in a specific direction. While this approach holds transformative potential, the ability for companies to have control over their cognitive architectures is equally critical. In this tutorial, we’ll explore the concept of cognitive architectures, the agent cognitive architecture, and the importance of control over cognitive architectures. We’ll also discuss how to create and manage cognitive architectures using LangChain’s tools, such as OpenGPTs.
Understanding Cognitive Architectures
In the context of LLM applications, cognitive architectures refer to the orchestration of an application’s reasoning processes based on contextual inputs and decision-making. The cognitive architecture of an application comprises two main components: how context is provided to the application and how the application reasons based on this context.
At LangChain, we’ve identified different levels of cognitive architectures that developers are building:
- Single LLM call: Determines the output of the application based on a single LLM invocation.
- Chain of LLM calls: Utilizes a sequence of LLM calls to determine the application’s output.
- LLM as a router: Uses an LLM to choose which action (tool, retriever, prompt) to employ.
- State machines: Utilizes LLMs to navigate between steps in a loop, with enumerated transition options in the code.
- Agents: Removes much of the scaffolding, allowing the transition options to be wholly determined by the LLM.
The Agent Cognitive Architecture
The Assistants API and GPTs represent examples of the “agent” cognitive architecture. This architecture revolves around using the LLM alone to define transition options, leading to a loop-based interaction model. In practice, applications using the agent architecture operate in a loop where the LLM is called based on user input, resulting in either a response to the user or actions to be taken.
The agent architecture empowers applications with an unconstrained model, enabling them to pull context as needed. While this approach is suitable for simple tasks, complex applications often require more control over the cognitive architecture.
Importance of Control Over Cognitive Architectures
Having control over the cognitive architecture offers several advantages, including enhanced flexibility and the ability to tailor the architecture to specific problem spaces. Furthermore, how context is provided to the agent plays a crucial role in the overall performance and reliability of the application. Effective cognitive architectures often involve a significant amount of pushed context, enabling developers to enforce the relevance of context to the LLM.
Introducing OpenGPTs by LangChain
LangChain provides tools to help developers create and manage cognitive architectures effectively. OpenGPTs, an open-source project, aims to recreate the experience offered by the Assistants API and GPTs, providing a configurable and editable version of these tools. OpenGPTs supports multiple model providers and allows easy modification of the retrieval methods used.
By leveraging tools like OpenGPTs, developers can gain more control over their cognitive architectures, tailor them to specific use cases, and ensure the reliability and performance of their applications.
In conclusion, while OpenAI’s bet on a cognitive architecture holds promise, the ability for companies to have control over their cognitive architectures is crucial for building reliable and high-performance applications. With LangChain’s tools like OpenGPTs, developers can take charge of their cognitive architectures and create transformative agent-like systems with enhanced flexibility and control.