Anthropic, the company behind the Claude AI models, recently shared a number that stopped a lot of people: the majority of its own code, more than 80% by its own count, is now written by AI rather than by its engineers. The company calls this progress toward "recursive self-improvement." In plain terms, that is AI helping to build the next AI.
It sounds like science fiction. For a Malaysian business owner deciding whether to hire a software team or wait for the robots, it raises a fair question: if AI can write its own code, why pay humans to write yours? Here is our honest read, minus the hype.
What does recursive self-improvement mean?
Recursive self-improvement is a feedback loop. A better AI model helps engineers write code faster, run more experiments, and train an even better model. That next model then speeds up the work again. Round and round.
Anthropic's own chart shows how steep the climb has been. Code shipped per engineer barely moved for years, then jumped sharply as its AI coding tools matured.
Code contributed per person, per quarter, annotated with Claude model releases. Source: Anthropic.
This is not a press release claim either. IEEE Spectrum reported the same thing in May 2026: Anthropic says most of its code is now written by its AI, with humans directing and verifying the work. So a mild version of "AI builds itself" is genuinely here.
What AI building AI does well, and where it stops
AI is very good at one thing: take a clear, well-defined task and produce working code for it fast. On the easy and routine jobs, the success rate is high and getting higher.
The interesting part is the tasks that are open-ended, where the goal is fuzzy and someone has to decide what "good" even looks like.
Success rate climbs across task types, but open-ended problems still trail the simple ones. Source: Anthropic.
That gap is the whole story. Anthropic even tracks a separate question: can the model pick a better next step than a human? It is improving, but it is the frontier, not a solved problem.
Judgment, deciding the better next step, is still the hard part. Source: Anthropic.
Writing code was never the bottleneck for most business software. Deciding what to build, why, and what trade-offs to accept was. AI just made the easy part cheap. It did not touch the hard part.
How this affects Malaysian businesses
We build software for Malaysian companies, and we use these exact tools every day. AI writing most of the code is not a threat to us. It is a discount we pass to clients. The same app that took six engineers now takes two or three, and the savings are real.
But notice what does not change. Your LHDN e-invoice integration still has to match MyInvois rules that AI does not know about your business. Your PDPA data handling still needs a human who understands Malaysian law. Your 12-year-old accounting system still needs someone to decide how it connects to anything new. And when a payment fails at 2am, a model does not answer the phone or carry the liability.
That last point matters most. You are not really buying code. You are buying judgment and accountability. AI generates plausible code quickly, which is also why reviewing AI-written code carefully has become more important, not less. Someone competent still has to own the result.
Is recursive self-improvement possible, or is it hype?
Both, honestly. A mild form is real and measurable right now. Jack Clark, an Anthropic co-founder, publicly put 60% odds on AI being able to autonomously build a better version of itself by the end of 2028. Smart people disagree loudly about whether that holds.
Our take: the runaway "AI designs superintelligence overnight" story is still speculation. What is not speculation is the boring, useful version already in your hands. Good software teams are getting faster and cheaper this year, full stop. The winners will not be the businesses that wait for the dramatic future. They will be the ones that hire a partner using these tools well today.
So do not panic about software houses disappearing, and do not believe anyone who says AI removed the need for human judgment. The judgment got more valuable, not less.
Three things to do about it
- Ask any vendor how they use AI. A team that cannot answer is leaving your money on the table. A team that says "we let AI write everything unchecked" is a risk. You want the middle: AI for speed, humans for review and accountability.
- Pay for outcomes, not lines of code. Now that code is cheap, judge a partner on whether the thing works, fits Malaysian rules, and survives contact with your real customers.
- Move on the cheap projects you shelved. Internal tools and integrations that were "too expensive" two years ago may now fit your budget. The cost floor dropped.
For a fuller checklist, see our guide on choosing a software development company in the AI era.
Thinking about how AI changes what you should build, or what it should cost? Let's chat. We will give you an honest take, no sales pitch.
