Problem

Thesis targets a bottleneck in AI and computational science: research remains dependent on serendipity and constrained by experimental throughput, while many scientific tools were built for a pre-AI world. Researchers therefore spend time on manual, fragile, error-prone workflows rather than on forming and testing ideas. The economic pain is slow, costly iteration; Thesis’s stated ambition is to make state-of-the-art AI discovery 10 times faster and 10 times cheaper.

The core use case is autonomous exploration of large experiment spaces: finding better machine-learning architectures and methods, then applying the same approach to difficult problems in biology and materials science. Instead of treating research as a sequence of isolated manual attempts, Thesis aims to turn it into a continuously improving search process.

Product / Service

Thesis is building Darwin, a “hill-climbing” machine that searches a vast combinatorial space of experiments. It uses the outcome of each experiment to optimize the next, creating a recursive loop in which model training, evaluation, and experimentation compound over time. Its public cloud product is positioned as a lab for ML research where users can run experiments, evolve architectures, and let agents conduct parts of the science in one workspace.

The delivery model is an accessible cloud platform: users can start for free, receive full platform access subject to daily token limits, and pay only for what they use without providing a credit card upfront. The intended benefit is to give researchers data-aware AI teammates that can monitor experiments, analyze results, surface relevant literature, and safely run background workflows, increasing research throughput while reducing the burden of repetitive experimentation.

Market

Thesis competes in the emerging market for autonomous AI research, AI-for-science, and machine-learning research automation. The closest public comparables are Sakana AI’s AI Scientist, which autonomously explores possibilities for scientific breakthroughs, and Google’s AI co-scientist, a multi-agent research collaborator designed to generate novel hypotheses. Thesis appears differentiated by emphasizing a closed-loop system for training models and evolving experiments, although the available evidence does not present a formal competitor list.

Traction is early but concrete: Thesis is a Y Combinator Fall 2025 company listed as active with a three-person team, and its LinkedIn presence said users could train models on GPUs through the platform in April 2026. The company reports state-of-the-art performance on OpenAI’s MLE-Bench, achieved in one month with $10,000 of compute, and offers a free-to-start, usage-based product. No revenue, ARR, or paying-customer figures are disclosed in the available materials, so Thesis is best characterized publicly as revenue-unproven and still in early commercialization rather than as a scaled software business.