Watching America Run Away With AI - Alistair Pullen (Cosine AI)

Value of the Information & Strength of the Argument

The value of the information is high, offering unique insights into the strategic and technical challenges of building a sovereign AI model in the UK. Pullen provides concrete examples and references to specific models and tools, grounding his arguments in practical experience. The argumentation is generally solid, with clear reasoning about the trade-offs between model size, active parameters, and inference costs. However, some claims are speculative, such as the estimated sizes of Anthropic’s models, and the discussion is naturally biased towards promoting Cosine’s approach. The argument that an inference company can compete with billions in funding is compelling but may be overly optimistic, as it relies on specific conditions and trade-offs.

Scientific Rigor, Source Quality, Title Accuracy

The discussion demonstrates a good level of scientific rigor, with references to specific papers (e.g., GRPO, GSPO) and tools (e.g., SWE-bench, ARC-AGI). The sources cited in the description are relevant and support the topics discussed. The title accurately reflects the content, focusing on the impact of US export controls and the strategic response from a UK lab. The speaker’s expertise is evident, but the lack of independent verification of some claims (e.g., model sizes) slightly reduces the overall reliability. The presence of a sponsored segment is noted but does not detract from the technical content.

The title accurately reflects the central theme of the conversation: the impact of US export controls on AI development and the strategic response from a UK-based lab.

8/10

The discussion is grounded in practical experience and references to specific models, tools, and papers. However, some claims are speculative and lack direct evidence, and the speaker's perspective is inherently biased as CEO of Cosine.

• 00:00The sovereign mandate and the Fable ban
• 04:02Millions vs billions: the inference-company model
• 07:19The consortium feedback loop
• 07:40Why open models lag the frontier
• 14:59MoE vs dense, and why active params matter
• 16:29Trajectories: the process-data advantage
• 19:48Beating slop: reward the process, not the answer
• 26:06Reusable abstractions and the epistemic wall
• 29:56Code review becomes runtime proof
• 37:32Do agentic harnesses still matter?
• 40:35Swarm: orchestrating hundreds of sub-agents
• 45:14Why memory is still unsolved
• 48:25Synthetic data and graders for RL
• 53:09The US export gift and supply-chain risk

• Cosine — Company website, referenced as the organization behind the sovereign AI initiative.
• Isambard-AI — UK supercomputer used for training the sovereign model.
• GRPO (DeepSeekMath) — Paper on Group Relative Policy Optimization, referenced in the context of RL training.
• GSPO — Paper on Group Sequence Policy Optimization, referenced as an alternative to GRPO.
• Incompressible Knowledge Probes, Bojie Li — Paper referenced in the description, likely related to model analysis.
• Estimating the Size of Claude Opus 4.5/4.6 — Blog post referenced for estimating model sizes, influential in architectural decisions.
• ARC-AGI (Francois Chollet) — Benchmark for AGI, referenced in discussion about model capabilities.
• SWE-bench — Benchmark for software engineering tasks, referenced in context of evaluation.
• SystemVerilog — Hardware description language, mentioned as a domain for coding agents.
• Colossus (xAI) — Supercomputer cluster, referenced in context of inference infrastructure.
• Mistral AI — European AI lab, compared in terms of model scale and approach.
• Anthropic — US AI lab, discussed in terms of model sizes and inference costs.
• Cohere — AI company, mentioned as an example of non-frontier models.
• DeepSeek — Chinese AI lab, discussed in context of open-weight models.
• GLM (Z.ai) — Chinese AI model, referenced as a recent competitive open-weight model.
• NVIDIA B300 — GPU hardware, referenced in context of inference requirements.
• gpt-oss-120b — Open-weight model from OpenAI, compared in terms of active parameters.
• Devstral 2 — Mistral's coding model, compared to gpt-oss-120b.
• Llama 70b — Meta's open-weight model, referenced as a dense architecture example.
• Claude Code — Anthropic's coding agent, discussed in context of trajectories and data.
• Swarm (Cosine) — Cosine's system for orchestrating sub-agents, discussed in the episode.
• OpenAI Codex — OpenAI's coding agent, mentioned in context of agentic harnesses.
• Lumen Outpost (Cosine) — Cosine's product, referenced in context of training and deployment.
• Kimi K2 (Moonshot) — Open-weight model, referenced in context of open-source capabilities.
• Andrej Karpathy — AI researcher, referenced in context of AI trends.

• Estimating the Size of Claude Opus 4.5/4.6 — The blog post aligns with Pullen's claims about model sizes and active parameters.
• GRPO (DeepSeekMath) — The paper supports the discussion on RL training methods.

• ARC-AGI (Francois Chollet)

• app.rescript.info
• notion.com