a16zCan Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert
At a glance
WHAT IT’S REALLY ABOUT
Why compute, data, and environments shape open-model catch-up dynamics
- Nathan Lambert argues frontier AI progress is driven by three forms of capital—compute, data, and talent—plus the organizational ability to deploy them effectively, which currently favors integrated closed labs like OpenAI and Anthropic.
- He characterizes Chinese model builders as talent-rich but compute-starved, with distillation and large-scale use of frontier model outputs acting as partial substitutes whose real impact is difficult to pin down in modern reasoning-era post-training.
- Lambert believes today’s biggest capability gains—especially outside verifiable domains like math/code—come from expensive, iterative data acquisition and RL environment building, creating an emerging industry with quality and reward-hacking risks.
- He predicts open and closed models will persist as tiers: closed labs monetize premium early markets while open-weight “near-frontier” models spread capabilities into privacy-sensitive and cost-optimized deployments via a slower, more fragmented ecosystem.
- To reduce fear and hype around recursive self-improvement, Lambert proposes that nonprofits like Trillium can publish open tooling, environments, and detailed automation “trajectories” that make frontier methods legible to researchers and policymakers.
IDEAS WORTH REMEMBERING
5 ideasModel leadership is mostly a capital-allocation and execution problem: compute, data, talent, and org design.
Lambert frames progress as an organization’s ability to marshal three scarce inputs: compute, data, and talent—plus the managerial/organizational capacity to deploy them effectively. He argues frontier labs (OpenAI/Anthropic/Google) currently lead because they’ve combined these inputs earliest and most aggressively, not because of a single “secret trick.”
China’s constraint is compute, not talent; distillation helps but is often overstated and hard to quantify.
He describes Chinese labs as “technically excellent” with comparable talent, but materially constrained by GPU access. Distillation and heavy use of model outputs can partially substitute for missing compute, but he cautions the real impact is uncertain and the public evidence suggests large error bars—especially for today’s reasoning-focused post-training.
The shift from SFT-heavy RLHF to RL/environment-heavy reasoning makes distillation less plug-and-play.
Lambert argues older post-training (SFT-on-teacher completions) mapped cleanly to distillation, but modern reasoning progress is more RL-heavy and environment-driven, making “just distill the best model” less straightforward. He suggests missing pieces may include mid-training steps and data/teacher quality that open builders can’t easily replicate.
Open and closed will coexist as economic tiers operating on different “exponential timescales.”
He predicts a tiered market: top closed models capture premium margins and reach high-value markets first (e.g., coding), while “near-frontier” models increasingly become open weights to reach different markets (privacy-sensitive, on-prem, lower-margin use cases). Open ecosystems move slower because revenue and user-feedback loops are fragmented across builders, hosts, and post-training vendors.
Frontier progress is increasingly about proprietary data and RL environments—especially in non-verifiable domains.
He claims many frontier gains outside fully verifiable domains come from buying/building high-quality data and environments where models can “act” and be hill-climbed via RL. This is capital intensive and can take years across sectors like legal, healthcare, and biology, contrasting with faster gains in verifiable areas like code and math.
WORDS WORTH SAVING
5 quotesWhen I think about model progress, it comes down to, like, different forms of capital that you put into them, and that is, like, compute data and talent.
— Nathan Lambert
To say that China is anything but equal on talent is gonna-- you're gonna have a hard time because they just have so many people being trained at such a high level. And what they don't have is compute.
— Nathan Lambert
People distill my products every single day. They take it and they sh- they strip it down to its raw parts. They test everything. That's called competition.
— Jensen Huang (quoted in conversation)
Fundamentally, like AI progress is about data and about writing better software to serve your models, and those are very different worlds.
— Nathan Lambert
The AI capability plot doesn't necessarily worry me. But the AI capability plot and the pretty much non-existing preparation of cross public cybersecurity and all of these things, like, there needs to be more preparation in the assumption that this is coming- than there is.
— Nathan Lambert
High quality AI-generated summary created from speaker-labeled transcript.
