Building Self-Accelerating AI with Mirendil

Benham and Harsh, both former research scientists at Anthropic, founded Mirendil after concluding that major frontier labs were structurally diverging from what it would take to accelerate all areas of science. Benham traces his conviction to the scaling law period at OpenAI, where he was part of the BlueShift team and later a founding member of BlueShift Labs at Google. Harsh cites Dario Amodei's essay Machines of Loving Grace as a key inspiration and started an automated pre-training project at Anthropic when the state-of-the-art model was Sonnet 3.7 and no one else wanted to work on it. Together they have been working on accelerating science with AI for approximately five years.

Mirendil's core thesis is that a company whose business model is training large models and charging for access is structurally disincentivized to share self-improving AI technology broadly. The founders left Anthropic specifically to rethink company structure, culture, and business model from scratch, with the goal of making self-accelerating AI available to everyone rather than restricting it. Harsh distinguishes several versions of self-accelerating AI, from AlphaGo-style self-play loops to systems that learn and improve their own capabilities in an unknown domain. Mirendil's specific target is systems capable of doing the work of an AI researcher or engineer, including writing low-level kernels and conducting research at high throughput, with the expectation that a sharp version of this leads to a broader system capable of conducting advanced science more generally.

Mirendil frames its self-accelerating system not as a single model recursively modifying itself but as an ecosystem of models and humans working together as one intelligent being. For the foreseeable future, perhaps a few years, humans will remain part of that ecosystem. Successive generations of AI make improvements to themselves, gradually reducing the need for human prompting over time, with the ultimate prompt for a sufficiently advanced version being simply to achieve goals. Jumps from Claude Sonnet 3.5 to 4.0 to 4.5 have materially extended how long an AI system can run on a problem before needing human oversight, and even tiny reductions in oversight requirements lead to large increases in token spend and effective outcomes. Mirendil reports accomplishing work with roughly ten times fewer people and resources than frontier labs typically require, and the founders' goal is to reduce the number of people needed to run a frontier AI lab from approximately 200 of the best people down to 10, then 2, then 1, then zero on the AI side.

Favorable agent scaling remains an unsolved problem. A tenfold increase in company size today yields only roughly 1.2 times productivity, and agent systems do not yet demonstrate favorable scaling in system-wide productivity either. The key unsolved problems are oversight, resource allocation, and prioritization of ideas across agents. Mirendil views itself as a small-scale experiment in achieving favorable agent scaling before scaling up. The competitive prize it identifies is time, with companies willing to pay ten times more compute to reach a milestone just one month ahead of competitors.

Coding is described as a cheat code for AI self-acceleration because much of the system is built on code, and self-acceleration is said to have already begun in small ways once coding models became truly useful. One important caveat is that continuing to scale pre-training is probably not sufficient to generate the net new knowledge needed to solve the kinds of scientific problems Mirendil is targeting, such as Alzheimer's disease prediction. Mirendil argues that models becoming as good as or better than the best scientists is insufficient on its own, because problems like Alzheimer's disease have so much structural complexity in terms of required data that existing models cannot be seen meaningfully accelerating progress within ten years at the current pace. Building an AI scientist through incremental improvements will not be enough, which makes speed of progress especially critical given that it is not even known whether solving Alzheimer's disease is possible or what fundamental limits exist.

On safety, Mirendil's Fable launch included guardrails restricting certain areas including bioweapons and AI research assistance, though the founders acknowledge that restricting AI research assistance is different in nature from restricting bioweapons synthesis. Anthropic's concern about AI research guardrails includes preventing adversarial states from using the technology to accelerate their own AI development. Mirendil's stated approach is to ensure each specific use case is positive-sum and safe rather than applying a broad blocking restriction, arguing that a focused company can get safety right where companies with broad product scope treat it as too much trouble.