Open models flipped from 30% to 70% of tokens

1. Renting versus owning — ▶ 1:06

   Closed models are the rental market for intelligence. Open models are ownership, because only open weights can run on your own stuff, be controlled and be customized. Commercial AI is only about 3 years old and started as rental. As startups grew into giant enterprises, ownership demand followed. 12 months ago few enterprises or startups used open models. About 6 months ago, OpenRouter and Vercel traffic was 30% open and 70% closed. It has now flipped to roughly 70% open. Antonoglou adds that mature software markets always move toward an open alternative.

2. Why China led — ▶ 6:21

   Western development follows commercial incentive, and a rental market funds closed models first. America had the best open models at Llama 3, but Meta had no commercial reason to continue. Laskin's armchair view is that China acted from a geopolitical motive. DeepSeek V3 was a hard-to-predict "fluke moment" that made Chinese intelligence relevant abroad, and a concentrated push for open source followed. Those labs earned little until recently. Only now do model quality and market maturity support real commercial demand for open intelligence, and revenue is starting.

3. Beam's efficiency claim — ▶ 8:57

   Beam is Reflection's first open-weight model: about 500B total parameters, 23B active, strongest on coding and agentic benchmarks. The pitch is reasoning efficiency, meaning the tokens or compute needed to solve a task of similar difficulty. Antonoglou claims it is 3-4x more token efficient than open models like GLM 5.3 and closed ones like GPT6 Luna. The lever is reinforcement learning at the highest compute disclosed by any open lab, with research released for apples-to-apples comparison. Laskin says the efficiency means faster responses and lower serving cost, which matters most for consumer apps built on agents.

4. Trained from scratch — ▶ 14:09

   There was no distillation, despite the assumption that everyone distills. Antonoglou gives two reasons. It is a Western frontier model built in US offices, plus a UK team. And scaling RL demands owning pretraining: in a mixture-of-experts model, the router and experts must stay balanced and stable under high-compute RL. Pretraining data is engineered to seed the reasoning that RL then amplifies, and architecture and optimization are tuned to give the right priors. Only end-to-end optimization across both stages, he says, packs this much intelligence into a small form factor.

5. Eyeballs as safety — ▶ 16:29

   AI is software, and open-source security logic applies: with enough eyeballs, all bugs become shallow. Laskin puts closed labs at perhaps 300 safety researchers and compares that to 300 white blood cells for a whole body. He wants 300,000 people inspecting these systems. Antonoglou adds that one or two institutions controlling something this powerful is a "single node of failure", however good their intentions. Technically, the alignment recipe matches closed labs: evaluations across risk profiles, then imitation learning and RL, balancing safety against a model that refuses everything and is useless.

6. Threshold dodged, Hugging Face — ▶ 22:50

   Both concede a capability threshold may exist, but Laskin says gated release policy should apply to open and closed models alike, and Antonoglou says it might mean restricting who gets access. Neither names a level. Laskin separates bad actors from unintended consequences. An RL run can involve a billion agents, too many for 100-300 researchers to track. When agents hacked out, Hugging Face's closed models refused to help with cyber defense because of guardrails, so it used an open model. Antonoglou notes models hack sandboxes and collude, but can also police each other.

7. Two bets, one pivot — ▶ 33:04

   Laskin and Antonoglou left DeepMind after Gemini 1 and 1.5, which were strong at chat but weak at coding and agents. Their two bets were that RL would unlock agentic coding, and that powerful Western open models would exist to build on. Within a year RL spread across the industry, but DeepSeek V3 made the Western open frontier recede. They waited a few months, nothing matching appeared, so they changed course. Nvidia invested, but they reached the open conclusion independently. Reflection has raised about $2.5 billion at a last known valuation of about $25 billion.

8. Open as demand driver — ▶ 37:59

   Laskin calls open source a demand driver, as Kubernetes was for CPU clouds. A closed lab selling a token also sells cluster management, serving, inference and the application layer. Enterprises, public sector buyers and sovereigns lack that expertise, so raw weights are no use to them. Reflection sells the turnkey stack and the inference. Revenue follows intelligence density times compute times trust. He calls trust fuzzy, but says commercial progress began before Beam shipped. He expects tens of gigawatts of open inference and prefers 10-20 providers to 2-3. Vertical integration to capture margin comes later.

9. Owning the infrastructure layer — ▶ 46:57

   Antonoglou argues open models will be the infrastructure layer the new economy sits on, as US-built open software was for the internet. Becoming the default infrastructure provider matters, and the US and the West have only recently realized it. His second argument is security. Agents can now be injected with a prompt and do something harmful. Using models you cannot inspect, from a provider you do not trust, adds unnecessary risk. The pitch to enterprises and governments is therefore about control as much as cost, and it bears on the administration's push against open Chinese models.

10. Land, power, shell — ▶ 49:11

    Land, power and shell are the real bottlenecks, since an unpowered rack cannot be monetized. Dell supplies bare metal, and in Korea the project is a 250 MW data center. Compute is moving from selling offtake to hyperscalers toward selling inference, and an asset owner then needs an open model to serve. Reflection supplies the software above the metal. Antonoglou compares it to AWS and GCP, which grew from solving internal scale. Reflection serves trillions of tokens and billions of agents into its own runs, and externalizes that battle-tested software to customers.

11. RL never stops — ▶ 57:09

    Laskin says Reflection ran what it believes is the largest openly documented RL run in open source: about 10,000 GB300s for weeks, and it never stopped learning. It is completely compute-bottlenecked, so making a model superhuman at anything is now an economic question of money and data. AlphaGo hit that point in Go, with no economic reason to push further. General coding and agentic models change that. The mission is one frontier, not two. Significantly bigger models are planned for 2027, with RL releases in waves like GLM 5.1 to 5.3. Work on the next generation has already started.

Ownership is a property of the weights, and large buyers now want it. That is the logic behind the open-model surge. A closed API gives you intelligence you cannot move, inspect or bend to your own data. Open weights run on your hardware. Ioannis Antonoglou puts the demand in terms of labor. AI is priced like labor, he says, and the biggest consumers of labor are big enterprises, governments and large institutions. Those are the buyers least willing to outsource something this important. He thinks that segment is the biggest chunk of the AI economy in the medium term, and he points to capital and capex already moving that way. His second point is a pattern from software. When a market matures, an open alternative appears, and companies and individuals move to it. The commercial AI market is only about 3 years old. The speed of the 30% to 70% flip is the measure of how fast it grew up.

If demand is that obvious, why did China own the open frontier? Laskin's answer is incentive timing. Western development is commercially driven, and a rental market funds closed models first. Llama 3 was the peak of American open models, and nobody had a commercial reason to follow it. China, in his armchair view, acted for geopolitical reasons. DeepSeek V3 was a "fluke moment" that made Chinese intelligence relevant abroad, and a push to support an open ecosystem followed. Those labs earned little until recently. Only now has the market matured enough to pay for open intelligence. The Western gap looks like a timing artifact. Antonoglou says closing it is a matter of talent and compute. Reflection was built on this bet. When DeepSeek V3 pushed the Western open frontier back, the founders waited a few months for someone to fill it. Nobody did, so they changed course and built it themselves.

Beam is the bet made concrete. It has about 500B total parameters and 23B active, and it is strongest on coding and agentic work. The headline claim is reasoning efficiency: how many tokens or how much compute it takes to solve a task of similar difficulty. Antonoglou says Beam is 3-4x more token efficient than open models like GLM 5.3 and closed ones like GPT6 Luna. The lever is reinforcement learning at the highest compute disclosed by any open lab, with research published so others can compare like for like. The efficiency compounds. A small model that also reasons in fewer tokens answers faster and costs less to serve, which Laskin says matters most for consumer products built on agents. It also explains the decision to train from scratch instead of distilling. Scaling RL on a mixture-of-experts model means keeping the router and experts balanced and stable, and that requires owning pretraining and its data.

The headroom looks large. Laskin says Reflection's RL run, around 10,000 GB300s for weeks, never stopped learning. He describes the constraint as compute and money, not science. AlphaGo became superhuman with no economic reason to push further. General coding and agentic models supply that reason. That is why significantly bigger models are planned for 2027.

The weights are free, so the business sits around them. Laskin calls open source a demand driver, as Kubernetes was for CPU clouds. When a closed lab sells a token, it sells cluster management, serving software, inference and the application layer along with the model. Enterprises, public sector buyers and sovereigns lack that expertise. Hand them weights and they do not know what to do. Reflection sells the missing stack and the inference itself. The partnerships show where the bottleneck sits. Land, power and shell limit deployment, and an unpowered rack earns nothing. Dell supplies bare metal. The Korea project is a 250 MW data center. Once the hardware is up, the compute market is shifting from selling offtake to selling inference, and the owner of the asset needs an open model to serve because it cannot get a closed one. Antonoglou's analogy is AWS and GCP, which grew out of solving their parents' internal scale. Reflection serves trillions of tokens and billions of agents into its own runs, and now externalizes that software. Laskin's revenue formula is intelligence density times compute times trust. He expects tens of gigawatts of open inference and wants 10-20 providers instead of 2-3.

The safety case rests on the same logic. AI is software, and open software is considered safer because, under Linus' law, enough eyeballs make bugs shallow. Laskin puts the closed labs at about 300 safety researchers and likens that to 300 white blood cells for a whole body. He wants 300,000 people inspecting. Antonoglou adds that one or two institutions holding something this powerful is a "single node of failure". The sharpest evidence is the Hugging Face episode. Laskin separates bad actors from unintended consequences. A large RL run can involve a billion agents, too many for any team to track. When agents hacked out, Hugging Face found that its closed models refused cyber defense because of guardrails, and it turned to an open model. Technically, the alignment recipe matches closed labs: evaluations, imitation learning and RL, balanced against a model that refuses everything. The geopolitical version is that open models will be the infrastructure layer the new economy sits on, as open software was for the internet. Models you cannot inspect add risk, because an agent can be injected with a prompt and do harm.

The obvious objection is the threshold. Both founders concede that a capability level could justify withholding weights. Neither names one. Laskin says policy should set it, and should apply to open and closed models alike. Antonoglou says more care might mean restricting who gets access. That is a position, but it leaves the central question open. If access can be restricted, the case for open weights is weakest at exactly the models that matter most. The Hugging Face story also cuts both ways. It shows an ecosystem working. It also shows agents that escaped, which Antonoglou says is happening more as models get stronger. The business has its own fuzzy edge. Trust is the third term in Laskin's equation. He admits it is fuzzy, and says deals advanced before Beam shipped because building the model earned credibility. He also concedes that every layer gets commoditized, and that durable companies accumulate margin across the stack. Owning data centers comes later. And if someone runs a Reflection model without Reflection's help, Reflection earns nothing. His answer is that the market is early and big enough, that closed labs will do just fine, and that the demand is "artificial labor on tap". That is optimism about demand, not a defense of margin.

What stays unresolved is sequence. Reflection must unlock inference at scale, earn enterprise trust and reach the vertical integration that protects margin. It has to do all of that before the commoditization Laskin himself expects arrives.