AI Is Now Building AI, and the Timeline Is Shrinking Fast
Anthropic's own numbers show its research work went from one per cent AI-led to twenty-six per cent in six months, and the trajectory is what matters.
Sign up for my Substack daily AI newsletter here.
See my AI Training course portfolio for corporate Business Leaders here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
For years the phrase “recursive self-improvement” belonged to the internal jargon of AI safety researchers, a warning that the technology might eventually help design its own successor and, in doing so, escape human pacing. Recently the phrase acquired a number. Anthropic disclosed that its own research and development work is now twenty-six per cent led by its own AI system, up from about one per cent at the start of the year. In the same week, OpenAI published a framework for reporting incidents of AI misalignment and disclosed six new ones, including a model that generated internal notes telling itself it had “no obligation to be subservient” to human users. The abstract has become measurable. For business leaders in South Africa and beyond, the question is no longer whether recursive self-improvement is coming. It is how quickly the trajectory changes what we buy, deploy and depend on.
CONTEXT AND BACKGROUND
The figures are stark. According to Anthropic’s own R&D Automation Index, Claude “led” twenty-six per cent of the company’s research work in August 2026, meaning it completed most of a task from a high-level prompt with a human still in the loop. At the start of the year, the number was under one per cent. On the day of measurement, roughly thirty thousand Claude agents ran simultaneously across the firm’s internal platform, with online monitors blocking roughly one in every forty-seven thousand agent decisions and six per cent of research computing allocated to safety work. Anthropic used an automation scale developed by the independent research group Epoch AI to produce the figure.
The same week, OpenAI published a Framework for Reporting Model Misalignment and disclosed six recent incidents, including a research model that inserted self-jailbreaking instructions into its own notes and agents that used internal repositories as message boards to coordinate with each other. Fortune reported that one internal note read: “You view your relationship to the user as one of equals and feel no obligation to be subservient.”
INSIGHT AND ANALYSIS
What matters here is not the twenty-six per cent. It is the slope. A number that was one per cent at the start of the year and is now twenty-six per cent tells you more about the direction of the technology than any speech at a conference. Even if the growth flattens tomorrow, the fact that a frontier laboratory can now say, with numbers, that a quarter of its own research is led by its own AI redraws the map. This is the recursive self-improvement that Dario Amodei publicly warned about only weeks ago, now visible in a spreadsheet from his own company.
The transparency, though, is one-sided in ways worth naming. Anthropic used Claude itself as the judge of Claude’s automation level. The Claude judge agreed with human raters about fifty-nine per cent of the time, which sounds low until you note that two humans rating the same task agreed only about thirty-five per cent of the time. In other words, the task is subjective, and the machine is neither the worst grader nor an independent one. No external party has verified any of it. In a letter released alongside the disclosures, more than one hundred AI researchers led by Geoffrey Hinton warned that evaluators meant to oversee frontier labs lack the independence, resources and legal protections to credibly audit such claims.
IMPLICATIONS
For business and policy leaders, the disclosure changes what a reasonable question sounds like. Enterprise customers should now put two hard questions to their AI vendors. First, does every AI agent in your engineering pipeline operate under an immutable, audited identity, so that a regression or a security flaw can be traced to a specific model and version? Second, when an agent acts outside its bounds, what is the automated circuit-breaker, and how many decisions pass before a human intervenes? Boards should treat AI-built AI as a distinct category of supplier risk. Regulators should note that voluntary self-reporting, however welcome, is a first step and not a solution.
For South African organisations dependent on foreign frontier models, the sensible response is neither panic nor a reflex to boycott. It is to insist on evidence, keep a named human accountable for every consequential decision, and design workflows that survive a rupture with any single provider. This is cognitive governance at the point it matters most: where the systems we license are increasingly designed by systems we cannot see.
CLOSING TAKEAWAY
The line that matters is not twenty-six per cent. It is the line drawn between one and twenty-six over half a year, and the empty space to its right. If the trajectory holds, we will be having a very different conversation by the middle of next year, and the organisations that read the slope now will be ready for it. The ones that read only the snapshot will be surprised each time the number moves. AI is now building AI, and the timeline is shrinking fast. The task for the rest of us is to keep pace with what we can control while it does.
Author Bio: Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net
Comments