Charlie O’Neill joins Theo Jaffee and Sophia Dew to argue for an ecosystem where open and closed AI models coexist. He sees closed labs advancing the frontier while adaptable open models serve economically useful specialized tasks. His assessment of converging scaling recipes acknowledges that architecture, efficient serving and other optimizations still matter.
Charlie O’Neill discusses Kimi K3 as an example of a high-capability open model and emphasizes the combination of price, control and specialization. He argues that a stronger starting point can reduce the post-training needed to address general weaknesses before adapting a model to a particular domain. These are his deployment expectations, not a guarantee that very large models are cheap or easy to run.
Charlie O’Neill highlights training in persistent environments with terminals, coding tasks and simulated workplace tools. He views learning to act, inspect results and recover from mistakes as a shift from producing answers toward completing work. His forecast of more multi-trillion-parameter open models remains a prediction, and the discussion does not independently benchmark competing systems.
Watch on YouTube
