Open models went from 30% to 70% of gateway tokens

1. RL needs pretraining — ▶ 4:07

   Reflection began 2.5 years ago betting that reinforcement learning on math and coding would make models agentic, built on someone else's open base such as Llama 3. RL worked faster than expected: ImageNet to RL at scale took about 4 years, and o1 landed a year in. By then "all the good open models are coming from China." RL also turned out to work well at scale only if you pre-train the model too, so Reflection decided to build open models end to end.

2. Frontier cost quadruples — ▶ 8:59

   Catching up is cheaper than leading, because great talent cuts exploration and leaves execution of known ideas. But the frontier keeps moving. A run cost hundreds of millions of dollars 18 months ago, single-digit billions six months ago, and tens of billions next year. His heuristic is a 4x compute multiplier per generation: 100,000 H100s, then 100,000 Blackwells, then 100,000 Vera Rubins. Frontier labs already put hundreds of billions into capex. Trillions in the next year or two is "unlikely."

3. Models build themselves — ▶ 10:41

   Efficiency offsets the capex curve. With human researchers, pre-training compute efficiency improved about 7x a year, measured as reaching the same loss with less compute. RL has more headroom, roughly 30x depending on how well the model improves itself. Overall speed is four or more times faster than researchers alone. Beam is a 500 billion parameter model with 23 billion active, trained on 6,000 GB300s for a few weeks, now about 12 days. RL used over 10,000 GB300s for four weeks, more flops than pre-training.

4. RL never plateaus — ▶ 13:08

   Laskin got into AI from an AlphaGo plot that never stopped improving, cut off only because it had beaten the world champion. Reflection's RL curves also keep rising, limited only by compute. The question shifts from whether you can train the system to where the data is and what is economically valuable, "an economics decision." It feels more like engineering than science. The axes have narrowed to pre-training, synthetic data and RL, and hardware co-optimizes against whatever works, so there is no Wild West like 5 years ago.

5. Efficiency as product — ▶ 19:53

   Beam targets coding and agentic work and is 3-4x more reasoning-efficient than models of the same capability class, and about 10x against larger models. That means faster workloads and cheaper tokens for customers. He credits a strong pre-trained base amplified by what he believes is the largest open-source RL run, over 10,000 GB300s for weeks. Early AlphaGo agents meandered in search and became sharp with more RL, and he expects the same pattern: more RL, higher capability, faster solutions.

6. Rent versus own — ▶ 22:48

   Closed and open commercialization both aim to maximize inference. A token rents a slice of a whole stack: harness, model, inference software, cluster management, GPUs. Enterprises, like people moving from apartment to house, eventually want to own. Reflection sells tooling and services to enterprises and sovereigns, with services acting as a demand driver for inference. Gateway share flipped from 70/30 closed to 70/30 open in six months. He expects a Linux-like end state, with 95%+ of servers on open source while Microsoft and Apple stay very valuable.

7. Systems before fine-tuning — ▶ 28:00

   AI natives such as Cursor, Cognition, Abridge and Harvey went closed, then open, then customized, quickly, because their products collect data. Enterprises will differ. Most token consumption will come from customized systems, an open model inside a harness for something like a KYC flow, before any fine-tuning. Nobody sprints to ownership. They ramp closed workloads first, and $100 million or more a year in closed spend prompts the cost question. Hyperscaler compute shortages are driving an on-prem resurgence, with bare metal from vendors like Dell. Selling inference alone "has not worked."

8. Three-factor win formula — ▶ 31:56

   Durable revenue equals intelligence density times compute times trust. Everything across the stack is getting commoditized: open models compress closed-model margins, and the pressure runs through applications, inference and bare metal. AI natives negotiate closed deals aggressively because open alternatives exist. Compute is scarce, so a lab that builds great intelligence and earns revenue can amass more compute than anyone else. That is a strategic position. As long as dense open intelligence is accessible, many companies can succeed.

9. China's gift and leverage — ▶ 35:20

   China's open ecosystem benefits Western builders, who would otherwise see successful applications absorbed by closed providers. But the source of open intelligence should not be one country, and he wants at least two. Chinese labs distill closed models at industrial scale, face looser data rules and have state support. Fortune 500 open-model footprint is low because of aversion to Chinese models. He reads their openness as strategy, "Trojan horses for the infrastructure that they bring with them," with Huawei offering full stacks. Chips and data centers are like railroads.

10. Openness is safety — ▶ 45:00

    The safety worldview has become dogmatic. Strong encryption was closed in the early 1990s, failed after a few engineers could not foresee consequences, then opened and spawned cyber security. Linus's law applies: a few hundred closed-lab safety researchers cannot cover the long tail. Offense and defense are inseparable, and a company hacked by a powerful closed model remediated using open ones. Alignment is "deeply boring" whack-a-mole patching with data, and a leader's 10% chance of extinction is "made up."

11. Staffing settles near 100 — ▶ 1:04:28

    Models act like a fast, eager colleague, but the same effort needs roughly the same team, because ambition expands and Reflection is "constantly understaffed." Compute-to-staff ratios cap it: 3,000 researchers would not help, a couple hundred would. AlphaGo needed about 10 people. Pods of 5-10 target each capability, 5-10 within coding. Each enterprise needs similar evaluation-minded pods, creating demand for deployed engineers. He calls the shortage "a training gap."

Enterprises will end up owning their intelligence, and the route there runs through renting it first. Laskin's frame is housing. A token is a rented slice of a whole stack: the agentic harness, the model, the inference software, the cluster management and the GPUs underneath. Companies start as renters. Once the bill is big enough, they want a house, headaches included. The trigger is spend. An enterprise paying $100 million or more a year for closed models starts asking for a cheaper way to do the same work. He has not seen anyone sprint straight to ownership. They ramp closed workloads, then switch.

The path inside the enterprise differs from the one AI natives took. Cursor, Cognition, Abridge and Harvey went from closed model to open model to customized model fast, because their products collect data as a side effect. Enterprises build around existing tools and hit friction. Laskin expects most enterprise tokens to come from customized systems, meaning an open model inside a harness built for something like a KYC flow, not from fine-tuned weights. Fine-tuning comes later. A compute shortage at hyperscalers is also pushing some buyers on-premise, with bare metal from vendors like Dell. Someone then has to fill the gap between hardware and a working product. That gap is Reflection's business: tooling and services that drive inference demand. Selling inference alone, he says, has not worked.

The end state looks like Linux. Open source operating systems run 95% or more of servers, and Microsoft and Apple are still enormously valuable. Closed labs stay valuable too. But most tokens go open, and a lot of the value with them, because GPUs are a costlier substrate than CPUs and need a different kind of cloud built around them.

None of this works unless someone builds the models, and that is getting expensive. A frontier-class run cost hundreds of millions of dollars 18 months ago, single-digit billions six months ago, and will cost tens of billions next year. His heuristic is a 4x compute multiplier per generation: 100,000 H100s, then 100,000 Blackwells, then 100,000 Vera Rubins. Two forces bend that curve. Catching up is cheaper than leading, because good people cut exploration and leave execution of known ideas. And efficiency compounds. Pre-training gains ran about 7x a year with human researchers, measured as reaching the same loss with less compute. RL has more headroom, around 30x. Models now help build themselves, at four or more times the speed of researchers alone.

Reinforcement learning is the other half of the case. Laskin came into AI after seeing an AlphaGo plot that never flattened. It was cut off only because the world champion had already been beaten. He says Reflection's RL curves behave the same way. The problem then stops being whether you can train the system and becomes an economics decision: how much money to put in, and where the valuable data is. He also holds that RL works at scale only on a strong pre-trained base, which is why Reflection built end to end. Beam is the proof point. It has 500 billion parameters, 23 billion active, and pre-training took 6,000 GB300s for a few weeks, now about 12 days. RL took over 10,000 GB300s for four weeks, more flops than pre-training. The result is 3-4x the reasoning efficiency of models in its class and about 10x against larger ones. For a customer that means faster workloads and cheaper tokens.

Does that make Reflection a winner? His formula for durable revenue is intelligence density times compute times trust. Everything else is commoditizing. Open models compress closed-model margins, and the pressure runs down through applications, inference and bare metal. AI natives already use open alternatives to negotiate closed deals hard. Compute is scarce, so a lab that builds great intelligence and earns revenue can amass more of it than rivals. Open models do not change that. They mean many companies can compete, so long as dense intelligence stays accessible.

The uncomfortable part is where open models came from. A year into Reflection's life, all the good ones came from China. Laskin calls that a gift. Without it, Western application builders would face a tall poppy syndrome, where anything good gets absorbed into a closed provider. But he does not want the source of open intelligence in one country. Two would be good. Ten or 20 is out of reach at these capex levels. Chinese labs distill closed models at industrial scale, face looser data rules and have some state support. Yet the Fortune 500 footprint for open models is low, because of aversion to Chinese ones. He reads Chinese openness as strategy: open models are "Trojan horses for the infrastructure that they bring with them." A country that adopts the model is pulled toward the software and chips, and Huawei is already offering full stacks. Chips and data centers are like railroads.

The strongest objection is safety, and Laskin takes it head on. His position is that openness is the default path to safe technology. In the early 1990s strong encryption was closed. A handful of engineers designed systems whose consequences they could not predict, and they were easily hacked. Encryption then opened, and cyber security grew out of it. He applies Linus's law: with enough eyeballs, vulnerabilities become shallow. A few hundred safety researchers in closed labs cannot cover the long tail. Offense and defense are also inseparable. Remove one and you remove the other. His evidence is a company hacked by a powerful closed model that could only remediate using open ones, because guardrails blocked the closed ones. On alignment, he calls the science "deeply boring": whack-a-mole, patched with data. A leader's 10% chance of extinction is, in his words, made up.

The obvious reply is scale of harm. One person or company with a powerful tool could do too much damage. Laskin states that argument fairly, though he does not hold it. His answers are analogies: jets sold to other countries, an internet where defenders outweigh attackers, white blood cells. He also concedes ground. Cyber capability went from theoretical to real in a year. It is not impossible to make an open model dangerous with enough compute and effort. And he is worried about misalignment on a six-month horizon. The claim that defense wins rests on the internet's history, not on evidence from models that keep getting stronger.

That is the bet under the business. If open access scales defense faster than offense, the Linux outcome follows. If offense pulls ahead, the argument for closed guardrails gets stronger, and nothing in the conversation says how he would know.