Lineage chart debunks "China only steals AI" narrative: US models now adopt DeepSeek's MLA
kevinnbass · x · 2026-11-22
A model lineage chart by @rasbt undermines the "China only steals American AI" talking point: Meta's LLaMA used RoPE from China's RoFormer, DeepSeek used Google's GQA, and 2026 US models like Nemotron 3 Super and Trinity Large directly adopted DeepSeek's Multi-Head Latent Attention and expert routing bias.
The takeaway: the modern LLM stack was built by open research on both sides of the Pacific, while closed labs and regulators try to pull up the ladder.
More from AGI Musings
- Math community's wholesale AI rejection wastes top human time, dev warns — cgarciae88 · 2026-11-22
- Dataset builder says most academics told her LLMs aren't real AI — and advised "don't" — BlancheMinerva · 2026-11-22
- DeepMind's Samuel Albanie opens a bet on how many Millennium Prize Problems get solved by Jan 2027 — SamuelAlbanie · 2026-11-22
- Consciousness claims that index on model intelligence: what would that imply about humans? — jzl86 · 2026-11-22
- Noam Brown's 2023 Tweet Resurfaces: General Reasoning Worth 1,000x Inference Cost — manubfr · 2026-11-22
- Anthropic's opening bell marks the end of the age when only humans could think — r0ck3t23 · 2026-11-22