There are prophets who disclosed that the research blueprint of Company A has been leaked, and Claude Code dates back to five years ago.

No way, you came back from the future??

Due to a copyright lawsuit, an internal research roadmap of Anthropic dated July 2021 was accidentally exposed.

It's very short, the PDF only has two or three pages of main content.

But after reading this document marked "Top Secret" at that time from start to finish, I was completely stunned (just kidding).

Today's Claude Code, Tool Use, AI Self-Judging, Test-time compute, even trillions of parameters and ultra-long context...

Five years ago, just a few months after the company spun off from OpenAI with only a dozen employees in total, all of these were clearly written down in black and white.

Roadmap from 5 Years Ago

Many people thought that the evolution of large language models from chatbots to code-writing and operating-system-running Agents was a natural progression as the technology matured in the past two years.

That's wrong.

At the very beginning of this 2021 roadmap, Anthropic set a very clear prototype positioning for its ultimate goal (Safe, beneficial AGI):

To build a general-purpose AI assistant that can interact with humans through natural language to create, edit and interpret software.

What is a general-purpose AI assistant?

It can answer general questions, write functions, find bugs, design APIs, and discuss code styles and best practices.

The most impressive part is the following sentence:

It needs to demonstrate the full capabilities of a software engineer colleague.

Wait a minute, isn't this exactly Claude Code??

At that time, not long after GPT-3 rose to fame, when everyone was marveling at AI's ability to write acrostic poems, Anthropic directly skipped the simple logic of automatic code completion, and aimed at a full-fledged software engineering agent.

Not only that, on the specific implementation path, there was also an extremely exciting internal debate.

Researcher Andy Jones raised a core divergence in his comment: should we first build a really easy-to-use programming assistant, or first prove the feasibility of tool use?

The team faced two paths at that time:

Practical path: Develop Repo Search, so that AI can access the type definitions of projects.

By the way, Tom Brown, the first author of the GPT-3 paper, also participated in the discussion on this direction, and suggested starting with code Repo Browsing.

Then comes the second path —

Research path: Develop printf debugging and editor operations, to first prove that the model can call tools.

Finally, CEO Dario made the final call:

In the short term, we will first verify the feasibility of tool use, and in the long term, we will evolve it into a super user-friendly programming assistant.

Looking back at the evolution process of Claude Code now, it completely follows Dario's words step by step back then.

Self-judging, Trillion Parameters and Ultra-Long Memory

Apart from the Coding Agent, this roadmap also includes predictions for underlying technologies.

For example, in the pre-training section, back in 2021, they clearly wrote that they would train large models with 1T (1 trillion parameters) and above, and explore very large contexts, or other mechanisms that can accommodate massive amounts of memory and experience.

In that era when 175B models were considered huge, they had already taken trillion-parameter models and ultra-long memory as standard infrastructure.

What's more impressive is the part about human feedback and Alignment.

The document clearly states:

Model self-judging.

Back in 2021, they were already researching how to let the model evaluate its own behavior and output quality, and how to take the model's own judgment as a feedback signal. This is exactly the core idea of today's AI Feedback and self-correction of reasoning models.

More interestingly, the document even mentions a concept that later became extremely important in large model research:

Test-time compute.

Of course, at that time, this research was placed in the context of alignment and discriminator research, and there was still a considerable distance from the later o1-style reasoning computing expansion.

But at least it shows that as early as 2021, Anthropic's research vision had already covered computing investment beyond model training.

All in all, Neo Lab is really capable, no wonder Ilya was so astonished.

I wonder what the cutting-edge labs today are going through every day.

Internal Chilling Moments

Of course, since it's Anthropic, talks about safety issues are inevitable.

But after reading it, I found that these people really love discussing safety, and they spent a very long section on it.

As a company established on safety, Anthropic foresaw the nightmares we are facing today when Agents are everywhere back in 2021.

Sam McCandlish, co-founder, raised a soul-stirring question in his comment:

How to safely deploy tools without sandbox isolation?

Dario immediately followed up and added:

Specifically, it refers to the Tool that executes code during deployment!

(ps: There is really an exclamation mark in the original text, your Dario is that passionate)

It's amazing.

When AI can really run code directly in the host environment, call Shell, and modify files, once it gets out of control or is attacked by prompt injection, the consequences are unimaginable.

Issues like Agent privilege escalation, environment escape, and data pollution that are being discussed all over the internet today, they were already discussing the sandbox boundary back in 2021.

Maybe it's because they wanted to take this path from the very beginning.

By the way, Amanda also appears in this report, she is the in-house philosopher of Anthropic that became popular some time ago.

She pointed out very seriously:

I am a little worried that we are doing alignment research on natural language, but we are mainly advancing model capabilities on code. These two types of research must always be closely linked.

If we do not conduct alignment research in the field where capabilities are being improved, the effectiveness of alignment testing will be limited. This may lead to ripple effects, for example, alignment research will directly become an isolated island because it was not integrated in the early stage.

She even directly proposed to establish a company principle:

Capability research and alignment research in various fields should be advanced synchronously in proportion, unless we can give sufficient reasons to explain why they should be separated.

One More Thing

Alas, I was wondering why Anthropic has always been one step ahead of OpenAI, it turns out that it's the first-mover advantage.

Anthropic has stayed focused for the past five years, relentlessly working on high-quality code signals, reinforcement learning and enterprise workflows;

While OpenAI has been fighting on multiple fronts all the way, doing text-to-image generation, Sora video, voice assistants, and search, which has more or less dispersed its energy.

However, when everyone is still reviewing the differences between the two companies' technical routes, a shocking scandal more outrageous than Hollywood movies suddenly broke out at OpenAI —

Guys, there's a mole inside OpenAI!!

According to the indictment released by US federal prosecutors, 35-year-old Abdulrahman Alfozan, a former engineer on OpenAI's product team, has been officially dismissed and faces federal criminal charges.

What did this guy do?

In the core conference room dedicated to executives at OpenAI's San Francisco headquarters, he quietly plugged a bugging device with advanced encryption and data transmission functions, wrapped in toilet paper, into the USB port behind the TV.

You know, this conference room is where executives like Sam Altman hold core closed-door meetings, where all the talks are about superintelligence R&D secrets, undisclosed product plans and government cooperation briefings.

And it got hacked that easily.

The world is really a ramshackle makeshift setup...

PDF:

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

This article is from WeChat official account "QbitAI", Author: Jay, published with authorization from 36Kr.