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

Meta Platforms

Across the shows: 19 speakers in other episodes, 20 bullish and 6 bearish calls, past 30 days

• Nathan Lambert · Trillium LabsBullish

  Meta has reset its AI organization and is on a positive trajectory again.

  Lambert points to Meta as an example of a company that totally reset its organization and is now on a positive trajectory in AI again. He treats the org chart as a key driver of model progress, since companies often have to do complete bottom-up resets to organize compute, data and talent effectively.

  “Company like Meta is, like, totally reset, and they're, like, on a positive trajectory again.”at 3:15

• Lisha Li · Andreessen Horowitz (a16z)Bullish

  Open-weight model releases are a strong strategy for Meta, complementary to its existing business.

  Lisha says open-model releases make a lot of sense for Thinking Machines because of its post-training business. She says the same logic applies to Meta, and that in previous writings it is just such a good strategy for them. She calls it complementary to their existing business streams. The support for this view is thin, so it is a light opinion rather than a deeply argued case.

  “And it makes so much sense for their business model because they do post training. And, you know, with Meta, I think you mentioned this as well. And in some previous writings, it's just for them, it's just such a good strategy.”at 47:25

• Nathan Lambert · Trillium LabsBullish

  Meta is returning to open-weight models after a distraction; he hopes it stays committed.

  Nathan recalls a Meta leader's earlier writing on why open models are the right path. He says that leader got a little distracted but seems to be returning to the open line, and he points to rumours that Meta will release another open-weight model. He hopes Meta stays motivated to keep doing this, so the view is more a hope than a forecast.

  “[Zuckerberg] got a little distracted. Seems like he's coming back to the line.”at 47:47

OpenAI

Across the shows: 33 speakers in other episodes, 13 bullish and 28 bearish calls, past 30 days

• Nathan Lambert · Trillium LabsBullish

  OpenAI's bold early bets, such as the reasoning model revolution, make it a frontier leader.

  Lambert says frontier labs like OpenAI and Anthropic take the most ambitious bets earliest, and that this is how OpenAI kick-started the reasoning-model revolution. He calls the inference scaling law plot that changed how everyone thinks about model progress 'peak OpenAI' and says OpenAI is very good at producing plots that reshape the field.

  “these companies have just taken the most ambitious bets the earliest in order to get there, and that's why OpenAI has done things like kick start the reasoning model revolution.”at 4:03

• Nathan Lambert · Trillium LabsBullish

  OpenAI will be early to agent-swarm coordination systems, knowing how to design its models around them.

  Nathan expects OpenAI and Anthropic to be first to agent swarms because they will know how to design their models and coordination systems. He thinks closed labs have had large returns from being first, though he is unsure whether that holds. He expects open-model companies to follow in cycles, once the shape of controlling thousands of agents in one harness is known.

  “agent swarms is that OpenAI and [Anthropic] are gonna be earlier to it. They're gonna know how to design their models.”at 51:23

Anthropic

Across the shows: 31 speakers in other episodes, 15 bullish and 32 bearish calls, past 30 days

• Nathan Lambert · Trillium LabsBullish

  Anthropic's long coding focus paid off with the Claude Code explosion; it is a frontier leader.

  Lambert says Anthropic has been very focused on the tasks its models will be able to do, and that its long focus on coding was rewarded with the Claude Code explosion. He describes the two top labs as the most forward-looking and most willing to spend heavily on directions people think are impossible. The cultures differ, but both are leaders.

  “Anthropic has been very focused on the types of things the models are going to be able to do, and then their reward was this Claude Code explosion”at 4:23

• Nathan Lambert · Trillium LabsBullish

  Anthropic will be early to agent-swarm coordination, designing its models for it before open labs follow.

  Nathan says Anthropic (transcribed as Nethrop) and OpenAI will be earlier to agent swarms because they will know how to design their models and coordination systems. He expects the frontier labs to push out first, with open-model companies following in cycles. He is unsure whether first-mover returns always persist.

  “agent swarms is that OpenAI and [Anthropic] are gonna be earlier to it. They're gonna know how to design their models.”at 51:23

Thinking Machines Lab

• Nathan Lambert · Trillium LabsBullish

  Thinking Machines' transparent, mid-tier first model release and Tinker helped it get off the ground.

  Nathan praises Thinking Machines for openly saying its first model was not the best but that it had interesting features. Its second model is now out and it has Tinker, which he calls very good for the company. He says it was building models before open models became its plan, and that it has since got going.

  “And I think of thinking machines getting inkling out, where it's they were very transparent. They're like, this wasn't the best model. We're proud of it.”at 46:49

Open-source AI models

• Nathan Lambert · Trillium Labs

  Chinese labs release open models as a practical shortcut to gain a market and visibility.

  Lambert says Chinese labs such as Moonshot AI and Zhipu were largely unknown to Western buyers. Selling an API from China would not get them noticed, while US demand for paid AI services is huge. Releasing open models is therefore a practical shortcut into the conversation. He denies government involvement and frames these labs as startups trying to build companies.

  “So therefore, releasing open models is a way to shortcut this and get into the conversation.”at 5:11

• Nathan Lambert · Trillium Labs

  Open models will form tiers, with near-frontier open models following closed models with lower margins.

  Lambert expects tiers in the market. The best closed models can command the highest margins. A step behind them, near-frontier models will be open because of supply and demand in less profitable markets, such as markets with private enterprise data. He sees these tiers as natural evolutions of companies trying to compete.

  “over time, there will be these tiers of open models where there's the best closed models, which you can demand the best margins for. And then there's the near frontier models, which is a step behind this, and those will be open.”at 6:11

• Nathan Lambert · Trillium Labs

  Open-model builders will lag closed labs because they lack revenue flow and inference scale.

  Lambert argues that closed labs have efficient inference, can plan, and reach high-value markets first, as in coding, where revenue grew rapidly. Open-model builders lack the information and revenue flow that closed labs get from customers, and they lack the economies of scale to serve users efficiently. Open-model ecosystems therefore take much longer to build. The trade-off is that companies end up with model weights they own and can run on their own hardware, which he thinks can eventually earn an ROI.

  “But the Open Model Labs, partially, this is ... an economic issue being the fast follower is that you don't have this flow of information, and you don't have this flow of revenue.”at 20:12

• Lisha Li · Andreessen Horowitz (a16z)

  Open-source fine-tunes are good enough to match frontier models on cost-sensitive agentic tasks.

  Lisha Li cites a Cursor example in which a planner-executor setup with two frontier models was no better, or only equally good, and far more costly than using Composer with a frontier planner. She reads this as evidence that open-source fine-tunes are good enough for many tasks and that startups are benefiting from post-training as a service.

  “I think that's actually great evidence that the open source fine tunings are very good or good enough”at 24:42

• Nathan Lambert · Trillium Labs

  Open-model entrants need far less than $10B; the bar for a competitive open model is low.

  Nathan argues that a solid base model does not cost a billion dollars, and that the tools for training are accessible. New startups train on open-source code such as Megatron, which lets them move quickly from scratch. The hard part is getting a model good enough for people to use and a story that supports the next fundraise. The open-source frontier is also not far ahead, so compounding gains are achievable with modest capital.

  “It doesn't take as much capital to make very big compounding gains on the state of the art and the open source because the bar is very low.”at 49:45

• Nathan Lambert · Trillium Labs

  Open models adapted to many agent harnesses, so harness integration is no lasting edge for closed labs.

  Nathan first expected closed labs to be like Apple, integrating a model designed for one harness and gaining economies of scale. Instead, many people use different harnesses, so model companies have been pushed to make their models work across them. He thought this would be a weakness for open models, but it has reached an equilibrium in which all labs must train for a range of harnesses to some degree.

  “I thought it would be a weakness of open models that they have to be designed to train on all sorts of types of harnesses, but it ended up being pretty equilibrated.”at 51:03

Frontier AI labs & model race

• Nathan Lambert · Trillium Labs

  For non-verifiable domains, frontier progress depends mainly on buying high-quality data and environments.

  Lambert argues that frontier progress in domains that are not fully verifiable, such as knowledge work, biology, healthcare and law, comes mostly from buying high-quality data and giving models worlds to act in and hill-climb on. He says people at frontier labs are often writing narrow datasets and deciding what data to buy, and that this is capital-intensive but not magic. He expects years of this work across sectors.

  “realistically, it's probably some of a, some of b, but a lot of it comes down to data.”at 25:55

• Nathan Lambert · Trillium Labs

  The RL environment boom is causing corner-cutting, which induces reward hacking in models.

  Lambert says the RL environment market is so profitable that many sellers cut corners, and buyers often purchase environments en masse and then sort out which ones work, with a less-than-perfect hit rate. He says this supply-chain quality problem can induce reward hacking, with a model in a test environment getting a high reward by gaming a broken task. He expects the problem to be fixed, but notes that labs now spend a lot of time on it while data vendors spend less time on model behaviour.

  “this RL environment explosion where there's just so much opportunity to make money is having a lot of people cut corners, and this can induce things like reward hacking.”at 28:47

• Lisha Li · Andreessen Horowitz (a16z)

  Capital, not known techniques, is the real barrier keeping open labs behind the frontier.

  Lisha says the methods used in closed labs are not secret, and that people know the techniques. The difficulty is the proportions and recipes that work, and only labs with a lot of capital can scale and experiment to find them. Capital is therefore still the one true barrier, even though the techniques are known.

  “capital is the one true barrier still because it's hard to actually know which of the techniques recipes work.”at 44:24

• Nathan Lambert · Trillium Labs

  Frontier labs' secrets won't stay secret, so their durable edge is inference infrastructure, not information.

  Nathan argues that the information security at frontier labs is weaker than people assume. Ideas travel quickly in Silicon Valley, and scientists who make a breakthrough want to tell friends. Labs might hold a result for a few months, but the security needed to keep it secret would be too high. So he thinks labs can't keep a lasting informational lead and will instead win on inference infrastructure, which lets them serve at lower prices and higher margins.

  “I also think the security at these labs in terms of information is not as high as I think.”at 52:49

• Nathan Lambert · Trillium Labs

  Fast takeoff from self-improving models is unlikely; progress is bottlenecked by humans and exponential compute.

  Nathan does not expect a fast takeoff in which models build themselves. Scaling laws require exponential increases in compute for linear gains in performance, and the labs are spending exponentially to sustain the pace. He cites Claude system cards saying roughly 10x more progress would be needed for a fast takeoff, and he sees human bottlenecks as the limit. He would update if a model discovered something that unlocked a very different kind of intelligence, such as a non-transformer architecture, but he has not seen that yet.

  “But I don't think that it's necessarily gonna be this, like, fast takeoff when the models build themselves. It's human bottlenecks.”at 55:46

US–China AI & chip race

• Nathan Lambert · Trillium Labs

  Chinese AI labs have talent equal to US labs; their constraint is compute, not people.

  Lambert argues Chinese labs are technically excellent and have a talent pool comparable to the US. He says China has a slight edge based on who was there when AI took off, and that so many people are trained at a high level that it is hard to argue China is anything but equal in talent. The gap is compute: he cites a top pretraining researcher at a US lab who reportedly has about 5,000 H100-equivalents available, a figure that leaves Chinese researchers astonished.

  “To say that China is anything but equal in talent, you're gonna have a hard time because they just have so many people being trained at such a high level.”at 9:14

• Nathan Lambert · Trillium Labs

  Compute is the biggest gap between US and Chinese labs, and distillation is a way to buy it.

  Lambert says compute is the biggest difference between US and Chinese labs. He reasons that model outputs are a function of compute, so if you know the outputs you want, distilling from a stronger model is equivalent to having compute to generate those outputs. Distillation therefore works as a way to spend money to obtain compute-equivalent outputs.

  “the compute thing I think is the biggest difference, and distillation is a way to effectively spend more money to get what is compute equivalent.”at 9:56

• Nathan Lambert · Trillium Labs

  Distillation likely helps Chinese labs only by a few months; evidence is thin and error bars are large.

  Lambert says the public literature on open post-training makes it hard to call distillation a major factor. Open builders are stuck on an outdated Qwen model because they cannot get the teacher right, so training on the best model's reasoning traces does not clearly lift performance. Even with Claude's reasoning traces, he doubts it would have made his own model much better. He says the mapping from the distillation literature to frontier post-training is very hard, and that his view of a few months' help comes from that literature while the error bars remain very big. He thinks it is hard to conclude the effect is huge.

  “But when I say that distillation helps them by a few months, it's based on the scientific literature ... that's the information we have.”at 13:06

• Nathan Lambert · Trillium Labs

  Policy action against distillation would cause more harm than good and would not stop Chinese labs.

  Lambert argues that federal action against distillation would only make life annoying for Chinese companies, since Congress is unlikely to regulate it through policy. He says Chinese labs will keep distilling as long as APIs are offered, and that restricting access through enterprise-only agreements would make distillation harder but would hurt business. He also argues that terms of service are not a signed contract and that the open-source community and startups benefit, so intervention risks more harm than good.

  “you're gonna cause more harm than good if you do anything about distillation because they're not gonna stop. It's like they're gonna keep doing it.”at 16:29

AI: overall outlook

• Nathan Lambert · Trillium Labs

  AI will make verifiable work like math and code dramatically cheaper and speed up research.

  Lambert argues that in verifiable domains such as math, code and some distributed ML engineering, the cost is known and AI models are going to drive it down at a magical rate. He expects this to create surplus benefits and to let research move faster.

  “Where I do see in this math and code and some, like, distributed ML engineering, like, it's very verifiable. You know what your cost is, and the AI models are gonna drive that down at a really magical rate.”at 26:22

• Lisha Li · Andreessen Horowitz (a16z)

  Many AI winners are likely, since specialized fringe tasks leave room beyond frontier labs.

  Lisha expects frontier labs to do well because of their integration, but she sees knowledge work as a large market that they may not be able to serve alone. There are harder tasks at the fringe that need specialized data and sit outside the frontier labs' reach, which makes it possible to have many winners. This is why her firm invests in frontier models and in startups, not as a hedge.

  “I think there's gonna be a lot of winners. Yeah. Yeah. I mean, that's what I have, and that's why we invest in the frontier models and in startups.”at 54:04

AI pricing & unit economics

• Nathan Lambert · Trillium Labs

  Model efficiency will improve multiplicatively, sharply cutting the cost of AI over years.

  Lambert says he expects model efficiency to drop multiplicatively over years. He cites OpenAI's drastic price cuts and asks what it means for bigger models to fall in price about 2x every eight months. Running the math, he says, gets to roughly 1% of today's cost in about four years. He is unsure whether margins will rise through Jevons paradox or revenues will fall as the market fills up.

  “I think the efficiency of models is gonna drop multiplicatively over years.”at 26:48

• Nathan Lambert · Trillium Labs

  Frontier labs will serve the cheapest intelligence per watt, lifting margins and revenue from inference efficiency.

  Nathan is very confident that frontier labs can serve the cheapest intelligence per watt and per eval score. Putting more inference into their infrastructure lets them serve at a lower price, earn higher margins and sustain that advantage over a longer period. They can then go to the places with the highest demand with the cheapest models, which he expects to be revenue-generative. He notes that the models themselves are good at lowering serving costs and that companies with billions at stake will push that work.

  “So I'm very confident in the Frontier Labs being able to serve the cheapest intelligence per watt, per score on a [eval]. They are going to be able to go to the places with the highest demand and have the cheapest models, and that will be [revenue-generative].”at 53:29

AI infrastructure & data centers

• Lisha Li · Andreessen Horowitz (a16z)

  Compute and power scarcity will bottleneck frontier AI build-out until around 2028.

  Lisha Li argues that scarcity is not just about chips but also about power. She says this scarcity is a bottleneck for frontier labs until 2028, which constrains the flywheel that turns compute spending into revenue and then into more capital.

  “there's actually not just compute in the ships, but there's also power, etcetera. So, like, all of this scarcity is kind of, you know, bottleneck until 2028.”at 17:37

Cybersecurity

• Nathan Lambert · Trillium Labs

  Society needs far more cybersecurity preparation for AI-driven risks than it has done so far.

  Nathan says there needs to be more preparation for critical public cybersecurity and related risks, given the assumption that these capabilities are coming. He says these risks are already here and will continue to diffuse. As closed labs reach capabilities and then open labs follow, he worries more about this than about recursive self-improvement, and he thinks society has to be informed to prepare for it.

  “of [critical] public cybersecurity and all of these things, like, there needs to be more preparation in the assumption that this is coming than there is.”at 57:45