Crusoe brings AI Talks back to San Francisco with a panel on model selection

Crusoe is bringing the next edition of Crusoe AI Talks to San Francisco on October 7. The series focuses on architectural decisions, not product launches. This time the centerpiece is the dilemma every AI team faces: call a closed frontier model, run and adapt an open one, or train something proprietary. Three speakers represent three different answers to that question: Jeffrey Morgan, CEO and co-founder of Ollama, which built the de facto tool for running open models locally; Michael Elabd, co-founder of Trajectory AI, which builds infrastructure for training and fine-tuning; and Erwan Menard, Crusoe's VP of product, whose company supplies compute infrastructure for both training and inference.

According to the organizers, there is no industry consensus, and that is exactly what makes the discussion relevant. Every path trades off cost, control, performance, and time-to-market differently, and the equation keeps shifting as closed and open models improve in parallel. The panel is expected to cover what each company runs in production today and how they got there — not theory, but accumulated experience.

The event starts at 17:00 with registration and drinks, the panel at 17:30, and a networking reception at 19:00. The series' official description emphasizes: "no product pitches, just a panel and an audience of AI leaders working through training and inference decisions, architectural choices, and strategic bets." The expectation is technical depth and honest opinions in an entrepreneurial atmosphere, aimed at developers and researchers who make daily infrastructure decisions, not C-suite executives.

The choice of Ollama and Trajectory as representatives of different approaches is deliberate. Ollama has become the de facto tool for running open models locally, with an emphasis on developer experience and deployment simplicity; Trajectory focuses on the other side of the equation — the tooling and infrastructure needed when a team decides to train or fine-tune a model for specific needs. Crusoe, as host and infrastructure provider, sits in the middle: it profits from both directions but also sees the data on what teams actually choose when their own money is on the table.

The open-versus-closed calculus is not settled. A year ago open models lagged significantly behind closed ones on most benchmarks; today the gap has narrowed across many tasks, and the cost of self-hosting has dropped. At the same time, closed models have added multimodal capabilities, longer context windows, and more aggressive API pricing. For a team building a product now, the choice is not binary — many organizations run hybrid: a closed model for critical, latency-sensitive tasks, an open model for high-volume or privacy-sensitive workloads. The October 7 panel is meant to reflect that mixed reality, not ideology.