@MaxForAI: If it weren't for the lawsuit, we would never have known that Anthropic had perfectly predicted the current trajectory …

If it weren’t for the lawsuit, we would never have known that Anthropic had perfectly predicted the current trajectory of AI development back in 2021.

I just finished reading this internal document from @AnthropicAI that was exposed in the litigation, called Concise Research Roadmap, with the final update in July 2021.

The document was written by Anthropic co-founder Jared Kaplan, and it still retains internal comments from a group of core researchers, including @DarioAmodei, @NotTomBrown, @AmandaAskell, @ch402, and others.

The document even bears the label “Highly Confidential–Attorneys’ Eyes Only.”

After reading it carefully, I genuinely felt a bit emotional.

Because the things we’re discussing today—Coding Agents, Tool Use, long contexts, RL, automated feedback, even Agent safety issues—are basically all already present in this research roadmap from five years ago.

And it’s far from just some vague concepts.

First, Anthropic had already explicitly proposed in 2021 the idea of building an AI assistant with the capabilities of a software engineer. On the first page of the document, they wrote that they hoped to develop a general-purpose AI assistant that could help users create, edit, and interpret software through conversation. This includes developing browser plugins, writing functions, discovering bugs, designing APIs, discussing code styles and best practices, and more. The ultimate goal was to enable the AI to exhibit the capabilities of a software engineer colleague. You’ll notice that this is exactly what Claude Code is doing today, right? Even more interestingly, they believed at the time that achieving this goal would require large-scale generative models, long-term memory, and reinforcement learning through automated rewards and human feedback. They even had the Roadmap mapped out very clearly.

Second, they specifically listed RL for Tool Use as a key research direction back then. The core of the research was to teach large language models to use tools, understand environments and goals, and continuously improve through reinforcement learning. The specific tasks they proposed at the time included search, printf debugging, and so on, and they were also preparing to study whether tool use could yield better scaling effects for models than supervised learning. A whole bunch of issues being discussed in the Agent field today were already on their minds five years ago.

Third, what surprised me the most was that they were even discussing the product form of Coding Agents back then. Tom Brown suggested that repository browsing was a simple yet valuable tool, and they could start from there. Nelson Elhage noted that if AI could proactively look up type definitions in projects, that would be a very powerful capability. Andy Jones went further and proposed two different paths. One was to build the best AI programmer, and the other was to first validate AI’s ability to use tools. Dario Amodei’s response was: Prioritize the latter in the short term, the former in the long term. And in the same document, Sam McCandlish was already worrying about how to safely deploy non-sandboxed tools. Dario particularly emphasized tools for executing code in the deployment phase. Now look at all the discussions today around sandboxes, permission controls, and code execution safety for Coding Agents. It’s really hard not to feel a sense of awe.

Fourth, they realized very early on that model capabilities and Alignment couldn’t be decoupled. Amanda Askell commented with concern that if capability research focused on the code domain while alignment research only addressed natural language, the future might see models with ever-stronger capabilities that alignment research couldn’t fully cover. So she stressed that the two should be tightly integrated. This is actually a key issue facing Agent safety research today as well.

I think the most fascinating part of this document is how it makes us rethink Anthropic as a company.

Back in 2021, ChatGPT hadn’t even been released yet, and the entire industry didn’t have nearly as clear an understanding of what large language models could do in the future as we do today.

But this group of researchers at Anthropic had already set the software engineering assistant as a concrete research goal on the path to AGI, and around model scale, data, long contexts, tool use, reinforcement learning, and safety, they had formulated a fairly complete research roadmap.

At least looking back from today, their judgment on the direction of technological development was indeed remarkably prescient.

And even more notably, many of the issues discussed in this document still haven’t been fully resolved even by 2026.

Five years ago, they were thinking about how to make AI a colleague to software engineers.

Five years later, we’re already starting to discuss how to enable AI to independently complete entire software engineering tasks, or even participate in improving its own models and R&D systems (AKA: RSI).

Sometimes AI development seems unbelievably fast, but when you look back carefully, you’ll find that some people actually knew exactly where they were headed a long time ago.

🤔

lmaoooo

是

It’s not really like that, I just happened to scroll past it in another group this morning, just bb you didn’t mention it before so I didn’t pay close attention (effort: low)

true

Some points were sparked by OpenAI, so I still think “prophecy” is fitting?

Yes, just established.