美国AI研究员的中国之旅：年轻人，追赶者，算力焦虑与“AGI展示厅” ｜专访Nathan Lambert【101视频播客】

Tech topics & takeaways (China trip + open-model ecosystem)

China vs. US open-source / “open-weight” model leadership

• Nathan argues the US ceded open-model leadership to China around summer 2025.
  • This is based on breakthrough releases from DeepSeek (V3/R1) and Moonshot/Kimi K2, followed by continued competitive momentum.
• On whether the US can catch up:
  • He believes Nvidia is the most plausible catalyst, since Nvidia benefits broadly from the success of open models.
  • He also notes that major US model labs (e.g., Meta/Microsoft/others) chose not to lean fully into open-weight in recent years.

Best China “open source” models (his current view)

• When pressed, Nathan names Kimi and ZAI as top picks “as of today” for best open models.
• He mentions potential candidates and timelines:
  • DeepSeek was the “unanimous king” in 2025, especially the V3 and R1 lines.
  • Xiaomi “might” take a crown later.
  • He notes that Qwen’s biggest models being closed-source limits adoption/mindshare at the very top sizes.

Why Chinese labs release lots of open models

• Framed as “path to relevance”:
  • Open releases help ensure developers/users try the models.
  • Otherwise, models can remain effectively invisible even if API access exists.
• Also framed as low-risk relative to their goals:
  • The strategy helps build an ecosystem → get feedback → acquire customers.
• Expectation: the “fast follow gap” between labs will continue.

Key constraint: compute scarcity

• Nathan’s main claim: the largest disadvantage for Chinese labs is compute/GPU availability.
  • He suggests US labs (Meta/OpenAI) likely have more GPUs than Chinese companies combined.
  • He cites a rumor example of individual US researchers having thousands of GPUs.
• Impact on strategy:
  • Labs design models around likely downstream usage and training/inference cluster shapes.
  • Smaller training clusters may impose a ceiling on scaling to very large parameter frontier models.
    • He discusses expectations around roughly 3–6T+ and implies further scaling becomes harder without the latest racks/chips.
  • Compute constraints push more work into efficiency.
    • He notes that innovations like DeepSeek-style memory saving may spread.
    • But US labs may sometimes decide incremental efficiency gains aren’t worth the engineering cost.

Domestic accelerators (Huawei chips)

• Consensus he heard: Huawei accelerators are strong for inference, not training.
• Many labs use Huawei chips for serving and ask Huawei to improve to win more customers.
• He expects this inference/training asymmetry to persist “for years”, with the caveat that outcomes depend on semiconductor progress and ecosystem support for running models.

Model quality gaps: benchmarks vs. real-world utility

• On traditional benchmarks:
  • Nathan estimates US models are ~6–9 months ahead.
• But he argues real-world “available performance” can be closer because Chinese labs:
  • Release sooner after internal RL/training cycles.
  • Example framing: an RL run might take ~a week, then the model can release “within a day.”
• He predicts:
  • The benchmark gap may stabilize.
  • The US can still pull ahead in intangible areas, such as:
    • robustness in “knowledge work”
    • early adoption of productized workflows (e.g., cloud coding)

Coding models vs. “ease of use”

• Why a very good Chinese coding model isn’t fully dominant yet:
  • The US has an advantage in product UX (e.g., Cursor + Claude/Codex workflows).
  • Closed labs get consumer usage data back into training (Cursor/Codex/ChatGPT feedback loops).
  • Open-weight labs may produce strong raw code, but usage data often doesn’t feed back as directly.

Open-weight + agent workflows / ecosystem effects

• Agents are an emerging theme:
  • agentic products across different price tiers
  • knowledge-work tooling
  • cheaper experimentation due to open models
• He describes OpenRouter as:
  • community-oriented
  • smaller than major inference providers (e.g., Fireworks/Together)
  • still important for open-model mindshare and co-branding effects.
• Mentions agent examples such as Hermes agent (as part of agent popularity dynamics).

Enterprise strategy differences: why US “doesn’t build open” like China

• He contrasts corporate instincts:
  • US consumer instinct: “just use an API key.”
  • China corporate instinct: “we have compute/talent; let’s build models and then specialize/fine-tune for internal agentic products,” while still sometimes releasing general models for ecosystem adoption.
• Pattern described:
  • Release a general model, then fine-tune internal specialized agents/models
  • Example references:
    • Meituan-like internal use cases
    • Xiaomi as intertwined with robotics/car-cloud interfaces

Alibaba cloud vs. US cloud / distribution model

• Nathan claims China cloud (e.g., Alibaba Cloud) tends to carry its own models.
• By contrast, US ecosystems (e.g., AWS/Bedrock-style patterns) resell/host many models.
• He suggests China lacks some US-style “infrastructure reseller/optimization” layers (e.g., CoreWeave/Nebulous-style equivalents).

Data industry weakness in China (surprising gap)

• He highlights that China’s purchasable external AI training data ecosystem is less developed.
• He contrasts:
  • US frontier labs buying data from vendors for evaluation improvements
  • China teams more often relying on in-house environments/data
• He argues this can limit top-tier “agentic knowledge work,” which usually requires expensive domain training data (healthcare/law/finance-like labor markets).

Security and alignment framing

• Chinese labs emphasize “security” more than ideological/political “alignment.”
• Nathan supports open models generally but argues diffusion should be paced (e.g., not instantly uploading newly breakthrough closed weights publicly).

Cultural/organizational observations affecting execution speed

• He argues Chinese labs function as fast followers due to:
  • heavy young talent pipeline
  • fewer “priors” and willingness to adopt new paradigms quickly (MOE, RL, agents, tool-use)
  • structured education and competitive gating (which he also suggests may reduce the likelihood of “true visionary” culture at the Ilya Sutskever scale)
• He reiterates compute bottlenecks and organizational coordination:
  • building language models is partly an organizational problem
  • success depends on breaking big tasks into many parallel contributions.

---

Named products / systems / model families referenced

• Kimi / Moonshot (including discussion related to “Attention Residuals”; mention of young talent and open footprint)
• ZAI
• DeepSeek (V3, R1; “V3/R1 lines” repeatedly referenced)
• Xiaomi / model releases
  • references include a 7B “Mimo model”
  • mentions “Flash V2” and “2.5”
• Qwen (open strategy at small-to-medium sizes; big models described as closed)
• Alibaba / ModelScope and broader Alibaba cloud context
• Meituan (agentic products; internal specialized models not necessarily released)
• OpenRouter and agent products such as Hermes agent
• Cursor (plus hosted training/inference workflows, including Cursor training a Kimi model in RL sandboxes and serving via Fireworks)
• Huawei chips (strong for inference; training limitations)

---

Main speakers / sources

• Primary speaker: Nathan Lambert (AI researcher; interview subject)
• Interview host/source: Valley 101 (host name explicitly cited; exact person not otherwise identified)
• Referenced external figures:
  • Dario Amodei (Anthropic CEO; discussed risk framing for open-weight models)
  • Jensen (Nvidia; discussed in relation to the China GPU debate)
  • Dark matter (referenced as an opponent-side figure in the GPU debate)
  • Elon Musk (reposted a Moonshot paper mention)
  • Kaifu Lee (01.AI; met during the trip; discussed the leadership/transition question)
  • Katherine Rentoul (trip organizer; mentioned once)