Cosine on Sovereign AI and Multi-Agent Systems
Machine Learning Street Talk · youtube · 2026-07-14
This interview explores why a country cannot outsource frontier AI. Using the UK export-restricted coding model Fable as an example, Cosine CEO Alistair Pullen explains why they built a "sovereign model" from scratch in the UK, trained on the Isambard supercomputer in Bristol.
He makes a business case: reasoning companies don't necessarily need billion-dollar training budgets. With millions in funding, national-level compute quotas, and industry collaboration feedback loops, it's possible to build competitive systems.
The second half focuses on agent engineering:
- Why open-weight models still lag behind the frontier in scale, active parameters, and data;
- Trade-offs between MoE and dense architectures, and how active params dictate model "feel";
- Reducing "slop" by rewarding the process rather than the final answer;
- Turning code review into runtime proof: letting agents genuinely reproduce and exploit bugs in a VM;
- How Cosine's Swarm system coordinates over a hundred sub-agents at once;
- Why memory remains an unsolved problem, and how synthetic graders enable RL for tasks without tests.
More from coding & agent
- Tenable and AWS launch a Black Hat build event for open-source security agents and MCP servers — Dave_Maynor · 2026-07-22
- Codex helps build Valdiluce, an open-world game with climbing, gliding and gondolas — Dimillian · 2026-07-22
- HeyGen adds a media-sourcing skill for coding agents with 75k images and 10k tracks — HeyGen · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- LangSmith adds tracing for Pipecat, LiveKit, OpenAI Realtime, and Gemini Live — LangChain · 2026-07-22
- An MCP server signs every AI agent tool call into a verifiable Merkle chain — Funky_Chicken_22 · 2026-07-22