Anthropic's Leaked Confidential Roadmap: The AI Industry's Playbook Was Written Five Years Ago

A copyright lawsuit unexpectedly exposed an AI company's hand. An Anthropic internal research roadmap called "Concise Research Roadmap," written five years ago, was made public due to legal proceedings. At that time, ChatGPT had not yet appeared, and the entire industry was still questioning whether large models were just a transitional product. Yet this document, only three or four pages long, had already listed Coding Agents, Tool Use, long context, RL, automated feedback, and even Agent safety issues.

One checklist, fulfilled over five years

This document was written by Anthropic co-founder Jared Kaplan and contains dense annotations from the company's core team. Tom Brown pointed out that "repo browsing is a simple, high-value entry point," while Nelson Elhage predicted that "automatically finding type definitions in a project" would be an extremely user-valuable feature. Facing the roadmap dispute between the "strongest coding assistant" and a "tool capability demo," CEO Dario Amodei set the direction on the spot: "In the short term, demo tool use first; in the long term, build the strongest coding assistant."

That year, Anthropic had only been founded for six months and had raised just $124 million in total. It was still 16 months before ChatGPT ignited the world and 20 months before the release of Claude 1. After reading the discussion, one netizen said bluntly: "So the Agents we're all racing to build now were their to-do list." Others pointed out that for frontier researchers these insights may not be surprising, but actually executing these directions without going off track is the real skill.

What they saw clearly in 2021

In the internal document, Anthropic explicitly wrote out the core path to AGI: build large generative models with long memory capabilities and fine-tune them through reinforcement learning that combines automated rewards with human feedback. All research projects revolved around this center.

They believed the first vehicle for testing this path was a "conversational software engineer assistant"—one that could create, edit, and explain software through dialogue, write functions, find bugs, design APIs, and ultimately reach the level of a software engineer colleague. This was precisely the prototype of today's Claude Code.

Around this goal, the team placed "scaling" first among all research directions and prepared from the start for trillion-parameter scale. Jared Kaplan was the first author of Scaling Laws theory, and he brought this understanding from OpenAI to Anthropic, thinking about how to train trillion-scale or even larger models. The document also discussed how to increase GPU utilization, adapt early to next-generation hardware such as TPU and Trainium, how to encapsulate complex engineering like block-sparse attention so that non-specialists could configure hyperparameters, and how to build security protection systems for core assets such as large model weights.

In pretraining and data, from the beginning they planned to train the industry's best-performing dual foundation model for language and code. The dataset covered not only natural language and code, but also diffs, PRs, debugging records, and evaluation results from code execution, fully covering the entire software development workflow.

Scaling Laws are the key to understanding the underlying logic of this roadmap. They refer to the predictable power-law relationship between model performance (such as loss value) and model parameters, data volume, and compute: on a log-log plot, performance declines linearly as scale grows. This means that as long as scale continues to increase, capability improvements can be obtained as expected without relying on accidental architectural breakthroughs. Jared Kaplan was one of the core authors of this theory during his time at OpenAI, and he brought the belief in "scale first" to Anthropic. This directly explains why the document put "scaling" first among all research directions and prepared in advance for trillion parameters, GPU utilization, and TPU/Trainium adaptation—these were not scattered engineering efforts, but systematic investments guided by Scaling Laws.

Block-Sparse Attention is a key engineering method for addressing the computational bottleneck of long context. The computation of standard attention grows quadratically with sequence length. When context expands to 100K or even 1M tokens, memory and compute become difficult to bear. Block-sparse attention splits the sequence into blocks and computes attention weights only between some blocks, reducing complexity to near linear. The document's mention of "encapsulating complex engineering like block-sparse attention so non-specialists can configure hyperparameters" shows that Anthropic already anticipated long context would be a must-win battleground and tried to productize low-level optimization to lower the barrier for researchers. This is consistent with the later route of Claude expanding its context window from 9K to 100K and then to 1M tokens.

Capabilities and alignment were bound together from the start

In the document comments, Amanda Askell anticipated the risk of code capability and alignment research becoming disconnected: code capability continues to strengthen while lacking alignment constraints, while alignment research depends on insufficient model capabilities that limit testing depth. Over time, alignment research would become an "information silo." She required capability research and alignment research in all areas to advance synchronously and proportionally. Safety could not wait until model capabilities matured and then be added on.

Interpretability was another long-term investment direction. Christopher Olah specifically broke down large models' neurons, attention heads, and internal functional circuits, hoping to understand from underlying mechanisms how arithmetic reasoning, multi-turn dialogue, and code generation are actually implemented inside neural networks.

RL for Tool Use was listed as a core research direction—teaching large language models to use tools, understand environments and goals, and continuously improve through reinforcement learning. Dario Amodei's vision was not limited to coding scenarios. He also proposed that natural language editors were worth exploring, such as letting models modify documents in a Google Docs-like environment.

Safety considerations ran throughout. Sam McCandlish asked at the time: "How do we safely deploy non-sandboxed tools?" Dario even warned that tools capable of directly executing code during deployment carried extremely high risk. Today's daily hot topics in the Agent field—sandbox isolation and permission interception—were all part of their reasoning back then.

One roadmap, realized over five years

Looking back five years later, Claude can already process 1M tokens of long text at once, and the model has iterated to the Opus series. Safety tuning evolved from human-labeled feedback to constitutional training where "AI supervises AI," and then to today's hybrid approaches.

At the end of 2022, ChatGPT exploded overnight with RLHF, but this method had obvious pain points: it required hiring large numbers of people for annotation, making it costly and hard to scale. Just half a month after ChatGPT's release, Anthropic published a paper proposing the "Constitutional AI" approach—first define a set of behavioral principles for the model, then let one model judge and revise another model's answers, iterating repeatedly. Internally, this solved the problem that safety tuning could not keep up with model capability growth. Externally, it defined Claude's polite, restrained product character.

In 2023, the first generation of Claude officially opened for external partnerships, focusing on long-text processing and safety attributes. In May, the context window expanded from 9K to 100K tokens; in September, the first version of the "Responsible Scaling Policy" was released, grading potential model risks. This was the industry's first systematic dynamic safety mechanism. In October of the same year, a funding round led by Google pushed Anthropic's valuation to $18.4 billion.

In 2024, Anthropic sparked a wave in interpretability. The team used sparse autoencoders to conduct an "internal scan" of millions of features inside Claude Sonnet, extracted 34 million features from the model, and located internal circuits corresponding to specific concepts such as the "Golden Gate Bridge" and "program vulnerabilities." They also ran an experiment: by manually activating the feature corresponding to the "Golden Gate Bridge," they made the model unable to avoid the Golden Gate Bridge no matter what question it answered. This was the first time humans achieved precise explanation and manipulation of internal mechanisms in a commercial-grade large model, meaning AI safety research no longer had to "open blind boxes."

In October of the same year, a new version of Claude launched a "computer operation" feature, allowing the model to look at screens, click a mouse, and type on a keyboard like a human. Just one month later, Anthropic open-sourced the MCP protocol, unifying how models connect with various tools and data. In the first quarter of 2025, OpenAI and Google successively announced support for MCP, and domestic players Alibaba, Tencent, Baidu, and Alipay also密集 launched related services within four months. This was the first time Anthropic made competitors follow it at the infrastructure layer.

In February 2025, Claude 3.7 and Claude Code were released together. The former can switch between "quick answer" and "deep thinking" modes, while the latter is a dedicated coding assistant that can browse an entire codebase, check type definitions, debug, commit code, and run tests. All the discussions in the 2021 document became reality: from "repo browsing is a good entry point" to the rhythm of "demo capabilities first, then build productivity tools," everything landed without deviation. By the 4.5 version in September, AI coding capability had reached the level of a junior engineer.

In 2026, Anthropic densely updated its entire 5.5-generation model series, releasing from flagship to lightweight models in sequence, with multimodal, deep thinking, and agent capabilities maturing simultaneously. MCP has become the de facto standard for global AI tools, and Anthropic truly transformed from "a product maker" into "a rule setter."

Constitutional AI is Anthropic's alternative to the scaling pain points of RLHF. Traditional RLHF requires large numbers of human annotators to rank model outputs by preference, which is costly and difficult to scale along with model capabilities. Constitutional AI works by first writing a set of explicit behavioral principles for the model (the "constitution"), then having the model self-critique and self-revise according to these principles, and then using these self-generated preference data to train a reward model. This is equivalent to using AI to supervise AI, freeing humans from item-by-item annotation so they only need to set principles. It both solves the problem that safety tuning cannot keep up with capability growth and shapes Claude's polite, restrained product personality. The article mentions that Anthropic published a related paper half a month after ChatGPT's release at the end of 2022, indicating that this idea had already matured internally.

Sparse Autoencoder (SAE) is the core tool behind the breakthrough in interpretability research in 2024. Neurons inside large models are polysemantic; a single neuron often responds to multiple unrelated concepts at the same time, making direct observation hard to interpret. Sparse autoencoders force sparsity in the intermediate layer, decomposing superimposed activation patterns into a large number of monosemantic features. Anthropic used this to extract 34 million features from Claude Sonnet and locate internal circuits corresponding to specific concepts such as the "Golden Gate Bridge" and "program vulnerabilities." Manually activating the "Golden Gate Bridge" feature can make the model unable to avoid the topic, proving that these features causally influence output. This moved AI safety research from behavioral observation into internal mechanism manipulation, no longer "opening blind boxes."

MCP (Model Context Protocol) is a protocol open-sourced by Anthropic in November 2024 to unify how models connect with external tools and data sources. Before MCP, every tool or database integration required custom interface development, causing severe ecosystem fragmentation. MCP defines a standardized communication specification that allows models to call files, databases, APIs, and other resources in a unified way. It was quickly adopted by OpenAI, Google, and domestic players such as Alibaba, Tencent, Baidu, and Alipay, becoming the de facto standard for global AI tools. This was the first time Anthropic made competitors follow at the infrastructure layer, marking its rise from model capability competition to a contest over rule-making power.

Why this roadmap is worth revisiting

Over five years, industry trends changed round after round, and countless companies shifted direction chasing hype, while Anthropic steadily advanced along the line drawn back then. The value of this document is not which technology directions it predicted, but the rare capability it demonstrates: while the industry was groping forward in the dark, someone had already seen the map clearly and then walked step by step to the endpoint.

From underlying compute engineering, model foundation, and core capabilities to safety systems and social governance, Anthropic built a complete AGI R&D loop with no obvious weak links. This leaked document proves that their current achievements are essentially the fulfillment of technologies that were planned long in advance.

Reference link: https://i.smitop.com/concise-research-roadmap.pdf