Anthropic's 2021 Internal Research Roadmap Exposed in Copyright Lawsuit

Anthropic's internal document "Concise Research Roadmap," last updated in July 2021, was exposed in a copyright lawsuit. The document, authored by co-founder Jared Kaplan, is marked "Highly Confidential, Lawyer's Eyes Only" and includes comments from core researchers such as Dario Amodei, Tom Brown, and Amanda Askell.

The first page of the document proposes the development of a general AI assistant to help users create, edit, and explain software through dialogue, including developing browser plugins, writing functions, discovering bugs, designing APIs, and discussing coding styles and best practices. The ultimate goal is for AI to exhibit capabilities akin to those of a software engineer colleague. The path to achieving this includes large-scale generative models, long-term memory, and reinforcement learning through automated rewards and human feedback.

The document lists "RL for Tool Use" as an important research direction, focusing on teaching large language models to use tools, understand environments and goals, and improve through reinforcement learning. Specific tasks include searching and printf debugging, and exploring whether tool use can achieve better scaling than supervised learning. Tom Brown suggested that browsing code repositories is a simple yet valuable starting point. Nelson Elhage noted that actively searching for type definitions in projects would be a strong capability. Andy Jones proposed two paths: to create the best AI programmer or to first validate the AI's ability to use tools. Dario Amodei responded that the short-term priority is the latter, while the long-term priority is the former.

Sam McCandlish expressed concerns in the document about how to safely deploy non-sandboxed tools. Dario Amodei emphasized tools that execute code during the deployment phase. Amanda Askell commented that if capability research focuses on code while alignment research only addresses natural language, there may be a situation where capabilities are enhanced but alignment coverage is insufficient, thus both should be closely integrated. The document also discusses directions such as ultra-long context, trillion-parameter models, data quality, and computational testing.

This roadmap was established before the release of ChatGPT. Funding and research resources have shifted from early safety and scaling explorations to software engineering assistants and reinforcement learning for tool use. Current products like Coding Agent, tool invocation, and long context are highly aligned with the specific tasks and product forms discussed in the document. The disconnect between safe deployment and capability alignment issues remains unresolved five years later.

Source: Public Information