Two researchers left Anthropic after barely a year and founded Mirendil, raising $200 million at a $1 billion valuation. The bet is clear: bring to market the self-improvement capability that the biggest AI labs currently keep for themselves.

The in-house secret turned into a product

Major research labs—Anthropic, OpenAI, DeepMind—share a conviction: the fastest way to build better AI is to use AI to train AI. This involves techniques like reinforcement learning from human feedback, synthetic data generation, and iterative fine-tuning, all aimed at making each model iteration produce richer, more aligned training data. These methods are seen as a competitive advantage and remain unavailable as commercial solutions.

Mirendil wants to fill that gap. It's not about selling a ready-made LLM, but about providing the intellectual—and likely software—infrastructure for organizations to kick-start a continuous improvement cycle on their own models. The idea is as simple as it is ambitious: what Anthropic and its peers do for themselves, Mirendil intends to offer to anyone with proprietary data and the will to build a domain-specific edge.

What self-improvement means for an enterprise

Implementing self-improvement processes is not a minor upgrade. It requires an architecture that can handle multiple parallel pipelines: inference to generate output, automated quality assessment, collection of preference signals (human or simulated), and a new training phase. In an on-premise or self-hosted setting, this multiplies computational load. A single GPU for inference is no longer enough; a cluster is needed that can alternate training and evaluation cycles without bottlenecks.

Organizations weighing this deployment path face trade-offs that AI-RADAR regularly analyzes in its on-premise LLM frameworks. The choice between cloud and local infrastructure hinges on three axes: data sovereignty, long-term cost predictability, and the ability to tailor hardware for specific workloads. Labs that develop such techniques operate on thousands of GPUs with high-bandwidth interconnects. Bringing even a fraction of that complexity in-house demands careful attention to every component of the stack, from model quantization to orchestration systems.

The sovereignty and intellectual property puzzle

The rise of startups like Mirendil is part of a broader trend: the market is looking for ways to replicate the internal capabilities of tech giants without giving up control over data. In regulated sectors—finance, healthcare, defense—training a model on sensitive data through external APIs is often a non-starter. A self-improvement system running entirely on owned infrastructure has the potential to unlock currently impractical use cases. However, it brings operational and integration costs rarely captured by a simple hardware TCO.

Mirendil's billion-dollar valuation after a seed round signals that investors and enterprises see a high-value niche. Success will depend on the ability to package complex recipes into manageable products that don't require an army of researchers. For companies, it marks the beginning of an era where AI is not just a service to consume but a production process to internalize.

Beyond the funding round: what it means for on-premise builders

The emergence of players offering self-improving techniques reshapes the landscape for those planning their own AI infrastructure. It's no longer just about picking a pre-trained model and putting it into production; the lifecycle extends and demands continuous feedback loops. This shifts the discussion from the compute power for initial training to the sustainability of a system that learns day by day.

AI-RADAR dedicates in-depth coverage to the architectural choices for such workloads, comparing serving solutions, quantization strategies, and latency profiles in self-hosted environments. The message for technology decision-makers is that self-improvement isn't simply bought with a funding round: it must be built on solid foundations of infrastructure, skills, and data governance. The game, in short, is just getting started.