Alistair Pullen says Cosine is using UK government-backed compute to build Lumen Sovereign, a frontier model intended for critical British sectors. Rather than operate a large consumer inference service, Cosine plans to license weights for secure deployment. That reduces the infrastructure burden and lets the company focus limited compute on training, data and a narrower set of enterprise requirements.
Alistair Pullen argues that model scale, active parameter count and post-training dataPost-training adapts a pretrained AI model to improve its behavior, capabilities, or alignment after the main training stage. help explain the gap between open and closed systems. For coding, he wants reinforcement learningReinforcement learning trains an AI system to choose actions using feedback about the results of its earlier choices. to reward the specific decisions that make a trajectory good or bad instead of assigning one score to every token. Runtime validationDeterministic software verification checks an output with a repeatable procedure that produces the same result for the same inputs and state. also matters: an agent should demonstrate that a feature works in a realistic environment, not merely produce a plausible diff or rewrite its own tests.
The discussion also covers multi-agent swarmsA multi-agent system contains multiple AI agents that interact, coordinate, divide work, or influence one another while pursuing tasks., code-review debt and memory. Cosine's swarm can divide a project among planners and workers, while file locks and change notifications reduce interference. Alistair Pullen still expects human reviewCode review examines proposed source changes to assess their correctness, clarity and suitability before or around integration. because abstractions, maintainability and intent remain difficult to verify. The sponsor segment is omitted from this summary.
Watch on YouTube
