Why a Nation Can't Outsource Its Frontier AI - Alistair Pullen (Cosine AI)

<p>This episode is sponsored by Notion. Learn more about Notion&#39;s Developer Platform today at https://notion.com/mlst</p><p><br></p><p>Britain&#39;s most capable coding model can&#39;t be exported, and that ban is the whole reason Cosine set out to build one from scratch. Alistair Pullen, CEO and co-founder of Cosine, sits down with Tim Scarfe to explain how a frontier system he calls Fable, locked behind US export controls, became the founding case for a UK sovereign model trained on the Isambard supercomputer in Bristol.</p><p><br></p><p>The bet underneath it is economic. Pullen argues that an inference company, rather than a training-first lab, doesn&#39;t need billions to compete: millions, a national compute allocation, and a consortium feedback loop can be enough. From there it gets into the machinery, why open-weight models still trail the frontier on size, active parameters and data, the mixture-of-experts versus dense trade-off and why active params dominate how a model actually feels, and the edge that real coding trajectories confer.</p><p><br></p><p>The back half is about making agents trustworthy. Pullen makes the case for beating &quot;slop&quot; by rewarding the process instead of the final answer, reframes code review as runtime proof (spin the bug up in a VM and force the agent to actually exploit it), and walks through Swarm, Cosine&#39;s system running hundreds of sub-agents in one shot. It ends on why memory is still an unsolved hack, how synthetic graders let you run RL on tasks with no built-in test, and why Pullen reads US export controls as an accidental gift, with a supply-chain sting in the tail.</p><p><br></p><p>---</p><p>TIMESTAMPS:</p><p>00:00:00 The sovereign mandate and the Fable ban</p><p>00:04:02 Millions vs billions: the inference-company model</p><p>00:07:19 The consortium feedback loop</p><p>00:07:40 Why open models lag the frontier</p><p>00:14:59 MoE vs dense, and why active params matter</p><p>00:16:29 Trajectories: the process-data advantage</p><p>00:19:48 Beating slop: reward the process, not the answer</p><p>00:26:06 Reusable abstractions and the epistemic wall</p><p>00:29:56 Code review becomes runtime proof</p><p>00:37:32 Do agentic harnesses still matter?</p><p>00:40:35 Swarm: orchestrating hundreds of sub-agents</p><p>00:45:14 Why memory is still unsolved</p><p>00:48:25 Synthetic data and graders for RL</p><p>00:53:09 The US export gift and supply-chain risk</p><p><br></p><p>---</p><p>REFERENCES:</p><p>organization:</p><p>[00:01:15] Cosine</p><p>https://cosine.sh</p><p>[00:04:14] Mistral AI</p><p>https://mistral.ai</p><p>[00:05:50] Anthropic</p><p>https://www.anthropic.com</p><p>[00:07:42] Cohere</p><p>https://cohere.com</p><p>[00:08:36] DeepSeek</p><p>https://www.deepseek.com</p><p>tool:</p><p>[00:02:52] Isambard-AI</p><p>https://isambard.ac.uk</p><p>[00:05:56] Colossus (xAI)</p><p>https://en.wikipedia.org/wiki/Colossus_(supercomputer)</p><p>[00:07:52] GLM (Z.ai)</p><p>https://z.ai</p><p>[00:11:52] NVIDIA B300</p><p>https://www.nvidia.com/en-us/data-center/dgx-b300/</p><p>[00:15:37] gpt-oss-120b</p><p>https://huggingface.co/openai/gpt-oss-120b</p><p>[00:15:52] Devstral 2</p><p>https://mistral.ai/news/devstral</p><p>[00:16:01] Llama 70b</p><p>https://www.llama.com</p><p>[00:17:05] Claude Code</p><p>https://www.anthropic.com/claude-code</p><p>[00:26:23] ARC-AGI (Francois Chollet)</p><p>https://arcprize.org</p><p>[00:40:38] Swarm (Cosine)</p><p>https://cosine.sh</p><p>[00:40:50] OpenAI Codex</p><p>https://github.com/openai/codex</p><p>[00:41:16] Lumen Outpost (Cosine)</p><p>https://cosine.sh</p><p>[00:41:18] Kimi K2 (Moonshot)</p><p>https://huggingface.co/moonshotai/Kimi-K2-Instruct</p><p>[00:49:55] SWE-bench</p><p>https://www.swebench.com</p><p>[00:52:40] SystemVerilog</p><p>https://en.wikipedia.org/wiki/SystemVerilog</p><p>person:</p><p>[00:23:40] Andrej Karpathy</p><p>https://karpathy.ai</p><p>paper:</p><p>[00:27:10] GRPO (DeepSeekMath)</p><p>https://arxiv.org/abs/2402.03300</p