The artificial intelligence industry is entering a new phase in which the contest is no longer simply about building larger models. It is increasingly about creating AI systems that can act independently, learn from experience and potentially contribute to the development of their successors.
Two recent developments capture this transition: Meta’s rapid consumer success with Muse and the emergence of Mirendil, a startup founded by former Anthropic researchers that is pursuing self-improving AI.
Meta introduced Muse on September 8 as a personal AI agent designed to do more than answer questions. Running through a dedicated secure virtual machine, Muse can perform tasks on a user’s behalf, including researching and booking travel, sending emails and organizing projects.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Meta describes the system as an agent capable of learning from conversations and becoming more useful over time. Its early adoption has been striking. By September 25, Sensor Tower estimated that Muse had surpassed 3.4 million downloads.
While other analytics companies reported figures ranging from roughly 2.3 million to 4.3 million. Muse also reached the top position on both the U.S. App Store and Google Play Store during the month. The significance extends beyond download numbers.
Muse represents Meta’s attempt to turn artificial intelligence into a consumer interface for action. Instead of opening separate applications, searching websites and completing individual tasks, users can increasingly delegate portions of that process to an AI agent.
Goldman Sachs analysts have described this movement as a shift from AI as backend infrastructure toward AI as a consumer-facing platform. That transition creates an enormous commercial opportunity, but also introduces a different category of risk.
An AI that merely generates text can be checked before its output is used. An agent that can browse, communicate, purchase, schedule or manipulate digital environments has greater operational consequences when it makes a mistake.
Muse’s rapid expansion therefore makes questions around permissions, privacy, cybersecurity and human oversight increasingly important. At the same time, the frontier is moving toward AI systems that participate in their own development.
Mirendil, founded by former Anthropic researchers, is reportedly in discussions to raise as much as $1 billion at a potential $5 billion valuation. The company previously raised a $200 million seed round at a $1 billion valuation, according to Bloomberg.
The idea of self-improving AI is no longer purely theoretical. Anthropic recently disclosed that Claude is involved in approximately 90% of its research and development work, with the model leading about 26% of that work.
Anthropic says these activities remain under human supervision, meaning the company has not handed over autonomous control of model development. Nevertheless, the trajectory matters.
If AI can increasingly write code, conduct experiments, evaluate results and contribute to the design of subsequent models, the economics of AI research could change dramatically. Development cycles could accelerate while the distinction between the tool and the researcher becomes increasingly blurred.
The combination of Muse’s consumer momentum and investment in self-improving AI illustrates the industry’s broader transformation. The next competitive advantage may not simply belong to whoever has the biggest model, but to whoever builds the most capable agentic system—and establishes the strongest mechanisms for controlling what that system can do.