Where AI’s power meets its limits
The two ideas in Clark’s title point to the tension at the heart of his argument. The “bitter lesson” suggests that general methods able to scale with greater computing power eventually outperform systems built around specialized human knowledge. The no-free-lunch theorem cautions that no single method is best for every problem; success depends on whether a method’s assumptions fit the task.
Clark sees truth in both. Each generation of AI can take on more work where it has abundant training data. Scientific research, however, happens at the edge of what is already known, where data are sparse and models have not been validated. That frontier is where the scientist belongs.
This boundary suggests a division of labor. People frame problems, exercise scientific judgment and decide what conclusions the evidence supports. AI handles more of the work between an idea and a result: preparing data, building computing environments, writing and debugging code, reproducing baselines, running experiments and documenting the work.
Clark describes the technology as “a mech suit for the researcher,” amplifying a scientist’s abilities while remaining under that scientist’s direction. Acceleration matters only if the results can be trusted. He will discuss how a carefully designed AI “harness” can preserve the data, code, model configurations, unsuccessful trials and decisions behind a result.
He will also explore validated “surrogate” models, faster and less expensive approximations of complex physical simulations, and how Bayesian optimization can select the right model for a particular task and computing budget. His goal is to shorten the path from an idea to a result that can withstand scrutiny.
Why Clark is asking these questions
Clark’s perspective comes from connecting mathematical optimization, production software, high-performance computing and scientific validation across a career that began at Oregon State.
He earned bachelor’s degrees in mathematics, physics and computational physics in 2008 and conducted undergraduate research in fluid dynamics with Malgorzata Peszynska, university distinguished professor of mathematics. Clark credits Peszynska and emeritus physics professor Rubin Landau with giving him his first opportunities to engage in research.
“One of the things I loved about Oregon State was that they let me explore as much as possible,” he said in a 2025 College of Science interview.
Clark carried that interdisciplinary approach into doctoral research at Cornell University. He later created and open-sourced the Metric Optimization Engine and co-founded SigOpt, which Intel acquired in 2020. At Intel, he led teams working in AI and high-performance computing.
Across those roles, Clark learned that a model can become very good at improving the measure it is given, even when that measure does not reflect the outcome people actually want. That insight led him to co-found Distributional and eventually to build Talaria, bringing together three strands of his career: making models better, making them work at scale and making their results worthy of trust.
Clark has written candidly about testing his own assumptions and changing direction when the evidence demanded it. That willingness gives the seminar a human dimension and brings the fourth-generation Beaver back to the place where mentors first trusted him to pursue difficult questions.
Bringing the questions back to Oregon State
The questions Clark is bringing back to Corvallis intersect with work underway across Oregon State. The university is strengthening its capabilities in AI, data science and research computing while pursuing advances in clean energy, robotics, integrated health and biotechnology, and climate science.
His path from university research to open-source software, startups and national-scale computing also illustrates how scientific ideas can move into wider use. Clark wants the tools themselves to reach widely, too. Talaria is currently in private beta, and he says it will be released as open-source software at NeurIPS 2026 and remain free for open scientific work.
“The real promise of AI is improving quality of life,” Clark wrote following a recent talk at the Agentic AI Summit at UC Berkeley.
On Oct. 5, he will invite the Oregon State community to consider what it will take to deliver on that promise: AI that gives researchers more time to think, more ideas to test and a stronger foundation for deciding what is true.
RSVP to attend Scott Clark’s seminar.
