Ontology wants to automate scientific discovery end-to-end, starting with AI post-training R&D

Ontology wants to automate scientific discovery end-to-end

Ron Arel's pitch for Ontology rests on a structural observation: drug discovery, materials science, and AI post-training all share the same underlying loop. You propose an experiment, get feedback, learn from it, and use that to propose the next one. If you can build a system that runs that loop autonomously and efficiently, the domain is secondary.

The company is starting with AI post-training because it's tractable enough to show results while being hard enough to matter. Arel describes the core challenge as scaling experiments without burning compute blindly. Training 15,000 models to find the best one isn't realistic, so the system has to learn to extract more signal from fewer runs, developing its own intermediate rewards rather than waiting for full ground-truth feedback each time.

“Our mission is to automate scientific discovery, and we're starting with AI R&D. Today we have some pretty exciting results regarding automated post-training of other language models. We really believe in not building copilots. What we are really trying to do here is build fully autonomous systems that are deployed in R&D environments and just run the entire loop autonomously perpetually.”

Compute and cost

Early on, Ontology burned through venture capital to run experiments. Now the company has its own cluster, its own infrastructure, and compute partners in place. Arel says he's comfortable committing "hundreds of thousands of dollars" to a run when the system has earned enough confidence to justify it — but getting to that discipline took time and investment in infrastructure before a single experiment ran.

The verifiability question

Arel largely accepts the argument that AI is moving fastest in verifiable domains, but frames the harder problem as one of evaluator dependency rather than verifiability per se. As experiments get more expensive — training large models, running clinical trials — a system can't query the ground-truth evaluator on every iteration. His thesis is that systems should be designed to progressively reduce that dependency, relying on intermediate signals more and more, until the evaluator is only needed at the end of a discovery cycle. That's a gradual path rather than a hard domain switch, and Ontology is building toward it from the computational end.

Two-year target

Arel's stated two-year goal is for Ontology's system to be deployable as infrastructure into any computational R&D environment. Give it access to the data and the ability to run experiments, put it in a well-structured environment where it gets clean signal, and it runs autonomously — cranking out discoveries, with humans checking the output rather than directing the process. He's explicit that the company is "quite far away from that right now," still monitoring deployments and validating results manually.

The team draws from DeepMind, Anthropic, and Factory AI, alongside academic researchers. Arel and his co-founder previously ran a nonprofit research group funded by the National Science Foundation, where they published early work on test-time scaling with language models. That lineage runs directly into Ontology's core bet: that agentic search over an experimental loop, applied at scale, is how you automate discovery.