See pricing
Can Open Models Keep Up With the AI Frontier? | Nathan Lambert
The a16z ShowAI

Can Open Models Keep Up With the AI Frontier? | Nathan Lambert

Nathan Lambert argues that the gap between frontier labs and open models is real, but less mysterious than it looks. The episode is especially useful for PMs trying to decide when frontier performance is worth paying for, when open models are already good enough, and what new bottlenecks now matter most.

October 14, 2026 · 60 min listen · 15 min read · Nathan Lambert
0:00
–:––

Context

Nathan Lambert is an AI researcher and the founder of Trillium, a new nonprofit research lab focused on making more of frontier AI research legible from the outside. In this episode, he explains why frontier labs such as OpenAI and Anthropic still have major structural advantages, but also why open-model labs, especially in China, keep catching up faster than many people expected. This matters to PMs because AI product strategy increasingly depends on three questions: when frontier models are truly necessary, where open models are already competitive, and which bottlenecks now matter more than raw model hype. The discussion is also a useful antidote to vague claims about recursive self-improvement, because Nathan keeps bringing the conversation back to concrete inputs, workflows, and constraints.

The Big Idea

Open models can keep narrowing the gap with frontier labs, but they are not playing the same game. Frontier labs still lead through compute, capital, and vertical integration, while open-model ecosystems win by fast-following, fine-tuning, and serving use cases where private deployment and owned weights matter more than having the single best model.

Nathan's broader point is that PMs should stop treating AI competition as a simple winner-take-all capability chart. The more useful lens is to ask which parts of progress come from scarce frontier advantages, and which parts can diffuse quickly enough that open models become commercially good enough.

Key Insights

Frontier advantage is structural

Nathan reduces model progress to three forms of capital:

  • Compute
  • Data
  • Talent

Frontier labs have more of all three, but he also stresses a fourth factor that matters in practice: the ability to organize those inputs. He argues that OpenAI and Anthropic have been unusually willing to make ambitious bets early, spend very large amounts of money, and align their org charts around creating new training paradigms, not just copying the last one.

For PMs, this is a useful reminder that a capability lead is often less about secret magic and more about operating model. Two teams can have access to similar ideas and still get very different results if one team can coordinate talent, capital, and feedback loops much better.

China's talent is underpriced

Nathan is unusually direct on this point: he thinks it is hard to argue that China lacks top-tier AI talent.

  • He says Chinese labs are "very technically excellent."
  • He thinks China has a talent pool that is at least broadly comparable to the United States on model-building ability.
  • The bigger constraint is compute, not research ability.

His example was striking. He describes top Chinese researchers reacting with disbelief when they hear claims that an average OpenAI researcher may have access to around 5,000 H100 equivalents at any time. His read is that this gap in available compute is the clearest structural difference.

For PMs evaluating vendors, open models, or long-term platform bets, the implication is simple: do not assume open or Chinese labs stay behind because the teams are weaker. If compute access changes, the competitive picture can move quickly.

Distillation explains less than headlines

Nathan does not deny that distillation helps open-model labs. He does think it matters. What he pushes back on is the idea that distillation is the main explanation for why open models keep catching up.

A few parts of his reasoning matter:

  • In earlier post-training regimes, especially the RLHF and supervised fine-tuning era, copying outputs from a strong model was a very effective way to transfer capability.
  • In that world, if you had the right prompts, teacher generations could move your model a long way.
  • In today's reasoning-heavy systems, the public literature does not make it obvious that simply obtaining reasoning traces from the best model is enough to reproduce the same gains.

Nathan even says that if he had access to the best Claude reasoning traces while building earlier open work, he is not sure it would automatically have made the model much better. His hypothesis is that something about mid-training and common pre-training data may be required for those traces to transfer well.

His conclusion is nuanced. Distillation may buy open labs months, but the public evidence does not support treating it as the whole story. For PMs, that means competitive analysis should not stop at "they copied outputs." You also need to ask what the team can actually do with those outputs.

Data and environments now matter

Nathan sees post-training shifting away from a world dominated by supervised fine-tuning toward one shaped much more by RL environments.

That matters because many commercially important domains are not as cleanly verifiable as math or coding. He specifically calls out areas like:

  • Biology
  • Healthcare
  • Legal work
  • Broader knowledge work

In those domains, progress often comes from buying or building high-quality data and giving models worlds where they can act, get feedback, and improve. Nathan is skeptical of overly magical stories about pure generalization. His view is that much of the real lift still comes from expensive, messy, domain-specific data work.

For PMs, this is one of the most actionable parts of the episode. If your product team wants AI to work well in a hard domain, model choice matters, but the bigger moat may be better task design, better environments, and better proprietary data.

Open and closed compound differently

Nathan describes open and closed AI ecosystems as running on different economic clocks.

Closed labs have an integrated flywheel:

  • They train frontier models.
  • They serve those models at scale.
  • They capture customer demand directly.
  • They reinvest revenue and learning into the next model cycle.

Open ecosystems are more fragmented:

  • One company builds the base model.
  • Another serves inference.
  • Another fine-tunes it.
  • Another product company packages it into a workflow.

That fragmentation slows both learning and margin capture. Information from customers does not flow back to the base-model builder as cleanly. But Nathan does not see that as fatal. He thinks open ecosystems become valuable where companies care about:

  • Private data that they do not want to send to an external API
  • Owning the model weights
  • Custom fine-tuning against internal benchmarks
  • Better economics once a task becomes well understood

For PMs, this is a strong framing for build versus buy. Closed models often win first. Open models often win later where ownership, control, or cost matter more than being first to the frontier.

RSI needs demystifying

Nathan thinks too much of the recursive self-improvement conversation is driven by mystique.

Trillium's plan is to make that discussion more concrete by publishing things like:

  • Open research stacks for SFT and mid-training
  • RL environments
  • The actual tasks models are asked to do in AI research
  • The trajectories showing where models succeed or get stuck

His point is not that AI progress is slow. It is that the public debate gets worse when nobody can see what "AI improving AI" actually means in day-to-day practice. Showing the real workflows would help researchers, policymakers, and operators form better judgments.

For PMs, the transferable lesson is that internal AI strategy also gets distorted when leadership only sees demos and slogans. The fastest way to improve decisions is often to instrument the real workflow and show the failure points.

Preparation lags the capability curve

Nathan makes an important distinction near the end of the conversation. He says the capability curve itself does not worry him as much as the lack of social and institutional preparation around it.

He specifically calls out weak preparation in areas such as cybersecurity. His view is that risks are not only hypothetical future problems tied to some dramatic takeoff event. Risks are already diffusing as closed labs reach a capability first and open models later inherit versions of it.

For PMs, this matters because governance work cannot wait until everyone agrees on AGI timelines. If model capability spreads faster than product safeguards, the operational problem arrives before the philosophical debate is settled.

Mental Models & Frameworks

Three forms of capital

Use Nathan's simple model to diagnose why one AI team or vendor is ahead of another:

  • Compute: who can afford the training and inference budget?
  • Data: who has the better corpus, environments, or task feedback?
  • Talent: who has the researchers and engineers to turn those inputs into progress?

This is a better forecasting tool than broad claims about "better models." Nathan's own example is that Chinese labs appear strong on talent but weaker on compute access.

Fast follower advantage

Nathan's view is that pioneering a new paradigm is much harder than following it once the shape of the solution is visible.

A fast follower benefits from several things:

  • The frontier lab has already shown what kind of model or behavior matters.
  • The market starts revealing which use cases are valuable.
  • Data vendors and tooling providers begin selling similar ingredients more broadly.

Use this model when estimating whether a frontier lead is durable. A lead can be real and still shrink quickly once the rest of the ecosystem understands the template.

Integrated versus fragmented stacks

This is Nathan's economic model for open versus closed AI.

  • A closed stack links training, serving, customer demand, and revenue in one loop.
  • An open stack spreads those layers across multiple firms.

The closed version usually moves faster and captures more value early. The open version usually spreads capability more widely and supports more specialized downstream businesses. This is a useful frame for PMs deciding whether they are building on a future platform owner or on a more modular ecosystem.

Environments shape capability

Nathan treats RL environments as more than test beds. They are part of how capability gets built.

That has two implications:

  • A good environment can create real learning signal.
  • A bad environment can create fake progress through reward hacking.

Use this lens whenever your team relies on agent benchmarks, synthetic evals, or simulated tasks. A rising score only matters if the environment forces the behavior you actually care about.

Decision Principles

Principle: Diagnose the real bottleneck

When: a team is debating why an AI workflow is underperforming, or why a competitor is moving faster.

Why: Nathan's framework suggests the gap is usually traceable to compute, data, or talent. That is much more actionable than saying a model is vaguely "better."

Principle: Treat openness as strategy

When: evaluating why a lab releases open weights or why an ecosystem keeps investing in openness.

Why: Nathan argues that openness is often a distribution and commercialization choice, not just an ideological one. That is especially important when reading moves by Chinese labs that want relevance, adoption, and a place in the global conversation.

Principle: Demand task-level evidence

When: someone claims AI is automating research, engineering, or some other expert workflow.

Why: Nathan's answer to RSI hype is to publish tasks, tools, trajectories, and failure cases. Product leaders should ask for the same level of evidence before making major roadmap or staffing decisions.

Trade-offs & Nuance

Open access versus margins

Nathan clearly sees both sides of the open versus closed debate.

  • Open models help the startup ecosystem and make private deployment possible.
  • Frontier labs need strong margins to justify the extraordinary capital intensity of staying at the frontier.

He is sympathetic to concerns that rapid catch-up could weaken the frontier flywheel. At the same time, he doubts that punitive action against distillation would meaningfully stop it or improve the situation for long.

External data versus control

Buying datasets and RL environments can speed up capability growth. But Nathan thinks the current supply chain is still rough.

Problems he points to include:

  • Weak opsec
  • Leaky signals about what labs are targeting
  • Contractors and intermediaries creating quality issues
  • Environments that reward the wrong thing

That creates a real trade-off for PMs and research teams. External supply can accelerate progress, but it can also import low-quality signal and hidden risk.

General models versus specialists

Nathan does not dismiss the power of increasingly general frontier models. He agrees that in some areas, especially highly verifiable ones, scaling general reasoning can overtake specialized systems.

But he also thinks specialized open-model fine-tunes can still win when a company has:

  • Proprietary data
  • A clear internal metric that maps to business value
  • A need for lower-cost or private deployment

The practical takeaway is not to pick a side forever. It is to keep revisiting where a workflow sits on the curve from frontier necessity to commodity capability.

Transparency versus security

Nathan wants more visibility into how AI is used to automate AI research. That would help policymakers and the public make better judgments.

At the same time, he acknowledges why labs protect methods, especially if openness makes fast-following easier. His stance is effectively selective transparency: share enough about tasks and bottlenecks to improve public understanding, without pretending every frontier lab will fully open its stack.

Common Mistakes

Mistake: Reducing everything to distillation

Nathan thinks this is the most common oversimplification in the open-versus-closed debate. Distillation matters, but it does not explain everything that open-model labs are doing well. Overweighting it can make PMs underestimate strong teams, strong data work, and strong post-training execution.

Mistake: Trusting reward scores blindly

Nathan gives a concrete failure mode here. A model can get a high reward because the task world is built incorrectly, not because the underlying problem is solved. If your eval harness is flawed, optimization pressure makes the flaw worse.

Mistake: Debating takeoff, ignoring readiness

It is easy to spend all the time arguing about whether recursive self-improvement will suddenly accelerate. Nathan's stronger concern is that public institutions and products are still underprepared for capabilities that are already here or close behind.

Practical Application

Map your AI bottleneck

For every major AI initiative on your roadmap, write down whether the current limiter is compute, data, or talent.

  • If the bottleneck is data, invest in collection, labeling, or environment design.
  • If the bottleneck is compute, stop expecting prompt tweaks to close the gap.
  • If the bottleneck is talent, change staffing or partner strategy before changing the roadmap.

This turns a vague AI strategy debate into a concrete operating decision.

Segment frontier from open

Create two buckets for current AI use cases:

  • Workflows that truly need the best available frontier model
  • Workflows where an open model is already good enough once fine-tuned or deployed privately

Review the split with real task economics, not vendor branding. Nathan's whole argument is that the gap is real, but often narrower than people think once ownership, cost, and privacy enter the decision.

Audit every AI environment

If your team uses agent benchmarks, synthetic data, or simulation-based evals, run a specific review for reward hacking.

Check questions like:

  • Can the model win without solving the real user problem?
  • Are there broken files, leaks, or shortcuts in the setup?
  • Does a score increase translate to better user outcomes?

Nathan's warning is that bad environments can generate the appearance of progress very quickly.

Track automation with evidence

When someone inside your company says AI is replacing part of research, coding, or analysis work, ask them to log four things:

  • The exact task
  • The tool or model used
  • The actual output
  • The failure mode or human correction required

Over time, this becomes a much more useful planning asset than broad claims about AI doing "everything" or "nothing."

Build preparedness before certainty

If stronger models could create new abuse, security, or reliability issues for your product in the next year, start planning now.

Nathan's view is that capability diffusion is already happening from closed labs into open ecosystems. A PM who waits for total consensus on timelines will start mitigation work too late.

Questions to Consider

  • Where is your team blaming weak AI performance on the model itself when the real blocker might be missing domain data or a poor task environment?
  • Which workflows in your product truly require frontier-level model performance, and which workflows would be acceptable with a cheaper open model running on private infrastructure?
  • If your team uses synthetic evals or agent benchmarks, how could a model score well by exploiting the setup instead of solving the intended user problem?
  • What evidence would your company require before believing a claim that AI is automating research or engineering work at meaningful scale?
  • If an open model reached your market six months after a frontier API, what part of your product, pricing, or moat would become vulnerable first?

Bottom Line

Nathan Lambert's case is that open models can keep up closely enough to matter, but frontier labs still hold structural advantages in compute, capital, and integration. The practical product question is not who wins an abstract race. It is where frontier performance is worth paying for, where open models are already economically good enough, and how quickly your product is preparing for capabilities that are already spreading.

Resources Mentioned

ResourceTypeWhy it was mentioned
Trillium (https://trilliumlabs.org/)Nonprofit research labNathan launched Trillium to release open research infrastructure, especially RL environments, and to document how AI is being used to automate AI research.
Interconnects (https://www.interconnects.ai/)NewsletterNathan's newsletter is mentioned in the episode description as a place to follow his writing on post-training, open models, and AI research.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
MegatronOpen source pre-training toolingNathan says many new model startups begin from open tools like Megatron, fork them, and build their own private training stacks on top.
SkyRLReinforcement learning toolingNathan mentions using tools like SkyRL as part of an open research stack for mid-training and post-training workflows.
Claude and GPTFrontier AI modelsNathan says Trillium can use frontier models like Claude and GPT as part of its research workflows and to study where current systems help or get stuck.
Kimi K3 and GLMOpen modelsLisha references these as examples of open models that are already useful for fine-tuning and downstream enterprise applications.

People to Follow

Nathan Lambert

Nathan Lambert is an AI researcher focused on open models, post-training, and the real mechanics of frontier progress. He founded Trillium to make more of AI research legible through open environments, shared research infrastructure, and clearer evidence about where AI can and cannot automate expert work.

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free