Cosine's CEO discusses how they secured government backing to build the UK's first sovereign LLM, overcoming compute constraints and competing with billion-dollar efforts.
要点 · TL;DR
Cosine 利用政府算力，以创业预算打造英国首个主权大语言模型。 Cosine builds UK's first sovereign LLM on a startup budget using government compute.
大规模后训练和强化学习比架构更能区分前沿模型。 Post-training and RL at scale differentiate frontier models more than architecture.
合成数据生成对于编码等领域的强化学习至关重要。 Synthetic data generation is essential for RL in coding and beyond.
核心观点 · Key points
在出口管制禁止Fable后，主权AI对英国至关重要。 Sovereign AI is critical for the UK after export controls banned Fable.
Cosine利用政府在Isambard上的算力分配构建英国主权LLM。 Cosine builds UK's sovereign LLM with government compute allocation on Isambard.
后训练和大规模强化学习是前沿模型的关键差异化因素。 Post-training and RL at scale are key differentiators for frontier models.
随着模型改进，智能体式框架的重要性正在下降。 Agentic harnesses are becoming less important as models improve.
合成数据生成对于编码及其他领域的强化学习至关重要。 Synthetic data generation is essential for RL in coding and beyond.
反共识 · Contrarian takes
Cosine通过授权技术而非出售token来与数十亿美元的公司竞争。 Cosine can compete with billions by licensing technology, not selling tokens.
开源模型落后是由于务实的架构选择，而非能力不足。 Open-source models lag due to pragmatic architecture choices, not lack of capability.
强化学习中的奖励黑客需要信用归因，而不仅仅是最终结果奖励。 Reward hacking in RL requires credit attribution, not just final outcome rewards.
所有记忆系统都是权宜之计；真正的记忆应在潜在空间中。 Memory systems are all hacks; true memory should be in latent space.
代码审查应转向运行时验证和漏洞测试。 Code review should shift to runtime validation and exploit testing.
美国出口管制无意中推动了英国主权AI的努力。 US export controls inadvertently boosted UK sovereign AI efforts.
本期章节 · Chapters（共 18）
0. 引言与Cosine背景Introduction and Cosine's Background
1. 主权AI与英国政府指令Sovereign AI and the UK Government Mandate
2. 与巨头竞争：效率与商业模式Competing with Billions: Efficiency and Business Model
3. 项目范围与权衡Project Scope and Trade-offs
4. 模型竞争力与架构Model Competitiveness and Architecture
5. 开源为何落后于闭源Why open source lags behind closed source
6. 关键因素：参数、活跃参数、数据Key factors: parameters, active parameters, data
7. Notion广告插播Notion ad break
8. MOE与密集模型及推理权衡MOE vs dense models and inference trade-offs
9. 数据工程与Anthropic的Claude Code优势Data engineering and Anthropic's advantage from Claude Code
10. 用户轨迹对训练的价值Value of user trajectories for training
11. RL轨迹中的信用分配Credit Attribution in RL Trajectories
12. 智能体工程与人类参与Agentic Engineering and Human Involvement
13. 理解债务与对齐问题Understanding debt and alignment issues
14. 群体架构与编排Swarm Architecture and Orchestration
15. 用于RL的合成数据生成Synthetic Data Generation for RL
16. 主权AI与供应链风险Sovereign AI and Supply Chain Risks
17. 对AI访问不平等的挫败感Frustration with AI access inequality