Cosine on Building Sovereign Frontier Models
Machine Learning Street Talk · rss · 2026-07-14
This interview explores why Cosine is building a "domestic/sovereign frontier AI." The core context: the UK's strongest coding model cannot be exported, prompting Cosine to build its own models and systems from scratch.
Key takeaways from the interview:
- The company believes that an inference-focused company doesn't necessarily need billions of dollars in training lab investments; a few million, national-level compute quotas, and coordinated feedback loops can also work
- A discussion on why open-weight models still lag behind frontier models, particularly regarding scale, active parameters, and data
- An explanation of the trade-offs between MoE vs dense, and why active params noticeably affect a model's "feel"
- The second half focuses on agent usability: rewarding the process rather than the outcome, turning code review into runtime verification, orchestrating hundreds of sub-agents using Swarm, and how memory remains an unsolved challenge
- Mentions of US export controls, supply chain risks, and how synthetic scorers can support RL tasks without built-in tests
More from Companies & People
- Investor argues Palantir-Nvidia partnership should slash Anthropic's IPO valuation — pdamodaran · 2026-09-11
- PyTorch Day Korea 2026 launches first offline conf, CFP closes Sept 13 — PyTorch · 2026-09-11
- Class action accuses Anthropic of overselling Claude subscriptions with deceptive usage multipliers — The Decoder · 2026-09-11
- AI post-training and evals jobs pay up to $850K, with median offers at $210K–$325K — FinanceYF5 · 2026-09-11
- Only a 4-day window: timeline casts doubt on OpenAI's independent math result claim — gleech · 2026-09-11
- 89% of firms use AI, only 6% see significant ROI — the busywork illusion — mikeflache · 2026-09-11