Descriptions:
Machine Learning Street Talk sits down with Alistair Pullen, CEO of Cosine AI, for a wide-ranging conversation about sovereign AI, frontier lab economics, and the technical challenges of training best-in-class coding agents. Cosine, a London-based lab founded in Hoxton, recently received a UK government mandate — backed by the government’s sovereign AI unit — to build what Pullen describes as the UK’s first sovereign LLM, with compute allocation on the Isambard supercomputer cluster in Bristol. Pullen explains how the Fable export control ban jolted the UK AI ecosystem into urgency, and how Cosine is positioned to pursue frontier model training despite operating at a fraction of the budget of labs like Anthropic, OpenAI, and Mistral.
The conversation gets technically substantive on reinforcement learning for coding agents — specifically the problem of “slop” (verbose, low-quality code that passes correctness tests), perverse incentives created by pure correctness rewards, and the unsolved challenge of credit attribution across long token trajectories. Pullen describes Cosine’s approach to finding the “high-entropy” decision points within a 256,000-token rollout that actually determined whether an agent succeeded or failed, and argues that better credit attribution would dramatically improve RL efficiency.
The interview also covers the declining importance of agentic harness frameworks as base models improve (“a model can basically do any task with bash only”), the compute advantage that Isambard allocation provides to a startup that would otherwise spend tens of millions on infrastructure, and the geopolitical stakes of building sovereign AI capability in the UK. Essential listening for anyone following European AI policy, sovereign compute infrastructure, or frontier RL research.
📺 Source: Machine Learning Street Talk · Published July 13, 2026
🏷️ Format: Interview
