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

Summary

Alistair Pullen explains how a banned frontier coding model became the catalyst for Cosine’s push to build a UK sovereign LLM, backed by the UK government’s sovereign AI unit and compute on Isambard in Bristol. He argues a startup can compete on frontier AI with millions rather than billions if it avoids being an inference-heavy API business and instead ships weights into customer environments. The conversation then turns technical: Pullen says frontier performance is driven by architecture, active parameter count, and data, with dense trillion-parameter models preferred in principle but MoE used because of inference constraints. He also lays out Cosine’s approach to training coding agents with trajectory-level credit assignment, additional reward shaping to avoid “slop,” runtime proof for code review, and swarm-style multi-agent orchestration. The episode closes on memory as an unsolved hack, synthetic graders for RL on tasks without clean tests, and the view that export controls may have inadvertently accelerated a more resilient UK supply chain.

Notes

• Cosine says it has shifted from regulated coding agents to building the UK’s first sovereign LLM.
• The sovereign AI effort is backed by the UK government’s sovereign AI unit.
• Cosine says the model will run on the Isambard cluster in Bristol.
• Compute is the main blocker for a startup-sized frontier effort.
• He says privately funding a project like this could require $50M-$100M of compute.
• Cosine is not an inference company.
• Customers usually run the model weights on their own GPUs or in their own hyperscaler accounts.
• Cosine licenses technology rather than making token margin.
• Skipping hosted inference reduces the company’s infrastructure burden.
• He thinks large labs are under pressure mainly because of inference, not training.
• There is little room for error because compute is constrained.
• He would prefer larger RL runs if more compute were available.
• He sees architecture, active parameter count, and data as the three core performance drivers.
• Mistral’s largest model is described as a 675B sparse MoE.
• He estimates Sonnet is roughly 1.1T-1.5T total params with 100B+ active.
• He estimates Opus around 1.5T-1.8T total params with 150B-180B active.
• DeepSeek V4 Pro is described as about 1.6T total with 30B-50B active.
• He ties those architecture choices to inference feasibility.
• He argues dense trillion-parameter models would be better if they were feasible to run.
• He says Devstral 2 feels better than GPT OSS 120B for coding.
• Cosine collects trajectories from its own usage and employee usage.
• He says trajectories mainly help reveal how users actually prompt models.
• Real prompts are messy, frustrated, and far less structured than RL datasets.
• Correctness-only RL can reinforce unsafe or ugly solutions.
• Cosine adds other rewards and sometimes reference implementations to shape behavior.
• He wants trajectory-level credit assignment, not equal weighting across a huge rollout.
• The goal is to reward the important decision points, not the filler.
• Coding is attractive because it has verifiable rewards.
• Non-verifiable domains like law are much harder.
• He thinks the practical route is to make tasks in-distribution and cover many cases at train time.
• Senior-quality maintainable software still needs humans today.
• Progress will come from model improvements, RL environments, and harness engineering.
• He says code review should become runtime proof.
• Cosine’s security product uses exploit validation in a VM to remove false positives.
• The same verification idea is being adapted to PRs.
• He says prove the happy path before a reviewer even reads the diff.
• He says agentic harnesses matter less as models improve.
• Cosine’s first agentic model was a fine-tuned GPT-4 Turbo in January 2024.
• Harness engineering still matters for token efficiency.
• Swarm is hierarchical multi-agent orchestration.
• Cosine used 253 sub-agents on a demo task.
• The demo was a mechanically constrained watch SDK/compiler with a 3D viewer.
• They use write locks and lower-level agent interjection to reduce agents stepping on each other.
• He still expects post-hoc human sweeping after swarm runs.
• Memory remains unsolved and current approaches feel like hacks.
• Vector DB and embedding memories are hard to query and hard to keep current.
• Memory is a risky RL surface because of reward hacking, leakage, and future-data access.
• Cosine is building synthetic graders from real completed work.
• They can target many languages and task types, including Java, Fortran, C, Verilog, and SystemVerilog.
• For Verilog and SystemVerilog they use a synthesizer or EDA-style flow to judge correctness.
• The UK already has the hardware needed for training, he says.
• He says geopolitical pressure has increased urgency from companies, government, China, and citizens.