x
The tech landscape is shifting on its axis. For years, we have treated artificial intelligence as an incredibly smart assistant — a chatbot that drafts emails, summarises documents, or suggests a neat block of code. But a profound transformation is happening behind the scenes. AI is no longer just a tool being built by humans; AI is actively building its own successors, and humanity might actually be slowing the process down.
A groundbreaking report by Anthropic, titled "When AI Builds Itself," reveals that frontier models have transitioned from passive assistants to active research collaborators. The data coming out of the San Francisco-based AI firm paints an astonishing picture of how quickly autonomous agents have taken over the heavy lifting of software engineering and machine learning research.
If you want to understand how fast this evolution is moving, you only need to look at Anthropic’s internal metrics. According to the company, its flagship model, Claude, now authors more than 80% of the code merged into Anthropic’s own codebase.
To put that into perspective, before the launch of the "Claude Code" research preview in early 2025, that figure sat in the low single digits. For the first four years of Anthropic's existence (2021–2024), the lines of code merged per engineer per day remained completely flat. The moment Claude was empowered to actually execute and run code — rather than just suggesting text for a human engineer to copy and paste — productivity skyrocketed.
Today, Anthropic’s engineers are shipping roughly eight times more code than they did in 2024. This eightfold acceleration means software development is moving at a velocity never before seen in human history.
This phenomenon is driving the industry toward a highly debated milestone: recursive self-improvement. This is the concept where an AI system becomes capable of fully autonomously designing, testing, and developing its own successor. Each new generation of AI builds a smarter version of the next, creating an exponential loop of intelligence.
While Anthropic explicitly states we are not fully there yet, the trajectory is clear. Current models are already running complex research experiments, identifying critical software vulnerabilities, and conducting sophisticated cybersecurity research.
This shift is not unique to Anthropic. The entire industry is racing toward highly autonomous, agentic workflows:
With deep-tech pioneers like Google DeepMind CEO Demis Hassabis predicting that Artificial General Intelligence (AGI) could arrive by the end of the decade, the window of time to prepare for autonomous AI systems is rapidly closing.
For decades, the primary constraint on AI development was compute power and algorithmic design. Today, the bottleneck has shifted. The biggest constraint on developing future AI systems may now be the human beings overseeing them.
As AI models take over coding and technical experimentation, humanity's role is shrinking into a supervisory capacity. Humans are no longer the builders; we are the validators. We are shifting our energy toward oversight, validation, and verification of an ever-expanding "virtual lab" run entirely by algorithms.
The biggest challenge now is not how to write the code, but deciding which problems are actually worth solving. Humans must provide the guardrails, the ethical boundaries, and the high-level judgment, even as our ability to comprehend the sheer volume and speed of AI-generated output is stretched to its absolute limit.
The implications of this shift stretch far beyond Silicon Valley tech companies. Anthropic points out that an AI system capable of automated research and development possesses skills that are easily transferable to other disciplines.
When an AI agent learns how to optimise its own code and run digital experiments efficiently, it can apply those identical methodologies to material sciences, structural engineering, genomic sequencing, and pharmaceutical discovery. We are on the verge of witnessing a revolution across all fields of science, driven by virtual autonomous labs that run 24/7 at a pace no human collective could ever match.
Recursive self-improvement is no longer a far-off science fiction trope. It is an unfolding reality, and it is arriving much faster than most global institutions, regulatory bodies, and societies are prepared for.
To read the full breakdown of this industry-shifting report, check out the original article on Decrypt:
👉 AI Is Already Developing AI, Says Anthropic—And Humans May Be Slowing Things Down
Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.
