AI Is Helping Build AI: What Anthropic's Latest Research Means for Future AI Engineers

Something interesting is happening inside one of the world''s leading AI labs.

AI isn''t only answering questions, generating images or helping developers write code anymore.

It''s increasingly becoming part of the process used to build better AI.

Anthropic recently published new measurements showing how Claude is being used across its own research and development work. According to the company, Claude led around 26% of the AI R&D work it measured in August.

AI was involved alongside humans in more than 90% of measured research work.

That sounds dramatic.

But before we jump to "AI is building itself," there''s an important distinction.

It isn''t.

Anthropic explicitly says Claude isn''t operating fully autonomously in the work it measured. Humans remain involved in the research process.

The more interesting story is what this tells us about the changing role of an engineer.

AI is moving deeper into the engineering workflow

Think about how software and AI development traditionally worked.

An engineer would research a problem, write code, run experiments, analyse results, debug failures and repeat the process.

AI tools initially entered that workflow as assistants.

They helped autocomplete code.

Then they started generating functions.

Then entire files.

Now AI systems can help with research, experimentation, analysis and larger engineering tasks.

Anthropic''s measurements give us a glimpse of what happens when AI becomes deeply integrated into that workflow rather than sitting beside it as a chatbot.

Reuters independently reported Anthropic''s figures and noted that Claude''s measured contribution has risen rapidly while remaining under human supervision.

The important change isn''t that engineers suddenly disappear.

It''s that the division of work between the engineer and the machine starts to move.

If execution becomes easier, judgment becomes more valuable

Imagine that an AI can generate ten possible solutions to an engineering problem in minutes.

Someone still needs to decide:

• Which problem is worth solving?
• What constraints matter?
• Which solution makes sense?
• What assumptions did the model make?
• How should the solution be tested?
• What happens when something fails?
• Can the output be trusted?
• Should the system be deployed?

Those aren''t simply prompting questions.

They''re engineering questions.

And they require understanding.

This is why students should be careful about building their entire AI skill set around whichever tool happens to be popular today.

Tools will change.

The ability to think through a problem lasts much longer.

The future AI engineer may do less manual execution

That doesn''t mean coding becomes irrelevant.

If anything, understanding code becomes particularly important when an AI is generating more of it for you.

You can''t effectively review something you don''t understand.

The same applies to machine learning.

If an AI produces a model with impressive-looking metrics, someone still needs to understand whether the dataset is appropriate, whether leakage occurred, whether the metric actually represents the business problem and whether the model will behave sensibly outside the training environment.

AI can accelerate execution.

It doesn''t automatically make the underlying decision correct.

So what should students actually learn?

Start with fundamentals.

Learn Python properly rather than knowing only enough syntax to prompt an assistant.

Understand data.

Learn how APIs work.

Understand Git and GitHub.

Learn the foundations of machine learning and deep learning.

Understand how modern LLM applications use retrieval, tools and external systems.

Learn how software gets deployed.

Then build.

Not tutorial projects where every step has already been decided for you.

Build projects where something goes wrong.

Work with messy data.

Connect an API that doesn''t behave exactly as expected.

Deploy something.

Debug it.

Document it.

Explain why you made the choices you made.

Those experiences develop something a prompt cannot give you instantly: judgment.

Learn to evaluate AI, not just use it

One of the most important skills emerging from AI-assisted engineering is evaluation.

Suppose an AI agent writes 500 lines of code.

The impressive part isn''t that you generated 500 lines quickly.

The important questions are:

• Does the code solve the right problem?
• Is it secure?
• Does it handle edge cases?
• Is it maintainable?
• Can another engineer understand it?
• Does it work in production?

Being able to ask those questions — and answer them — is much more valuable than simply generating more output.

AI should become leverage, not a replacement for understanding

There''s a tempting shortcut available to students today.

Whenever you encounter something difficult, ask AI.

Sometimes that''s exactly the right thing to do.

But if AI removes every moment where you would otherwise struggle, investigate and eventually understand something, it can also remove part of the learning process.

Use AI aggressively.

But don''t outsource your understanding.

Ask it to explain.

Ask it to challenge your solution.

Use it to explore alternatives.

Then verify what it tells you.

Build something with the answer.

Break it.

Fix it.

That''s where learning happens.

What Anthropic''s numbers really tell us

Claude leading 26% of Anthropic''s measured AI R&D work doesn''t tell us that human AI researchers are becoming unnecessary.

It tells us something more useful.

Human-AI collaboration is becoming deeply embedded inside advanced technical work.

Students preparing for that environment shouldn''t compete with AI simply by trying to type faster or produce more output.

They should become good at the things surrounding execution:

• Problem formulation.
• Architecture.
• Experimentation.
• Verification.
• Communication.
• Critical thinking.
• Engineering judgment.

And, most importantly, understanding the problem well enough to know when the AI is wrong.

The takeaway

A few years ago, learning AI meant understanding how to build models.

Then it increasingly meant knowing how to use powerful AI models.

The next stage may be about something broader:

Knowing how to work with AI systems to solve difficult problems reliably.

So don''t stop learning to code.

Don''t stop learning mathematics.

Don''t stop learning machine learning.

And don''t replace fundamentals with prompts.

Learn the fundamentals well enough that AI becomes leverage.

Because access to powerful AI will increasingly become common.

Knowing what to do with it won''t.

Want to build practical AI engineering skills? Explore CAIEE →