Is the Future of Frontier AI Open or Closed? Misha Laskin on Reflection AI’s Beam Model

Origins and the Bet on Reinforcement Learning

Misha Laskin frames Reflection AI’s mission as building frontier open intelligence and making it widely accessible. He describes the past year as a sprint to stand up a lab capable of doing this, comparing it to assembling a plane while flying it. The company started around 30 people and now sits near 300, with the team organized across pre-training, mid-training, and reinforcement learning. Their first open model, Beam, is a 500 billion parameter model with 23B active parameters.

Laskin’s answer to the question of what has been hardest is that “everything has been hard.” He argues that success depends on getting roughly 30 things right at once: talent acquisition and retention, culture, data, compute, and the infrastructure that makes compute usable. Coming from DeepMind, he notes that many tools he took for granted as a researcher worked only because years of buildout at a large lab had created a stable surface.

The founding thesis came from his time on the Gemini reinforcement learning team alongside co-founder Ioannis Antonoglou. At that point, Gemini 1 and 1.5 were primarily chat models, with roughly 95% of compute going to pre-training. The team had seen RL align chat models and believed the same approach applied to domains like mathematics and coding could produce agentic systems. They also saw open models emerging (a small Mistral model, Llama 2, and an expected Llama 3) and thought an open base could be built upon, enabling research to be done more capital-efficiently. Progress in RL arrived faster than expected, partly driven by o1. By about a year in, Laskin concluded that the best open models were coming from China and that the West lacked a strong open base for the technology they were building, so the most impactful move was to build open models end-to-end. He cites enterprise and geopolitical reasons, but emphasizes a research reason: pre-training is necessary to make RL work well at scale, so the two are tightly coupled.

Compute Scaling, Efficiency, and the Question of Asymptotes

On the economics of frontier training, Laskin distinguishes between exploration and execution. Catching up to the frontier is far cheaper than defining it, because execution of known ideas can be done with less capital than open-ended research. He estimates that a year or eighteen months ago a competitive frontier run was on the order of hundreds of millions of dollars, that six months ago it was billions, and that next year it will be in the range of ten billion. He roughly models each generation as a 4x compute multiplier, tying this to hardware progressions such as the move from 100,000 H100s to 100,000 Blackwells and then Vera Rubins.

When pressed on whether this scaling must eventually plateau given the revenue ceiling, Laskin agrees that capex is bounded but argues that efficiency gains offset the pressure. He describes pre-training efficiency improving around 7x per year when human researchers did the work, and says reinforcement learning has far more headroom, perhaps on the order of 30x per year depending on the model’s ability to improve itself. The models increasingly help build themselves, so progress accelerates. He frames intelligence extracted per training flop, and the revenue each flop can generate, as both rising, so the asymptote is not imminent.

Concrete numbers for Beam: pre-training ran on 6,000 GB300s for a few weeks, now reducible to about 12 days with infrastructure improvements. Reinforcement learning used a little over 10,000 GB300s for four weeks, so more flops were actually spent on RL, given its inference-heavy, sandboxed agent workloads. Laskin links his personal motivation to AlphaGo, where the RL curve never stopped improving and was cut off only because the champion had already been beaten. His point is that the field is now at a stage where RL systems are general and do not stop improving, so the question shifts from whether training works to an economic decision about how much compute to spend on capability gains. He characterizes the work as closer to engineering than to a new science, in the way rocketry is an engineering discipline with scientific content. He notes that the field has narrowed since the more diverse bets of early DeepMind and OpenAI, partly due to a hardware lottery in which hardware co-optimizes around what already works.

Where Value Comes From: Code, Agents, and Enterprise Verticals

Laskin says code and agentic capability form the foundation for general intelligence, and that generalization is surprising but jagged. His model adapted quickly to new data for particular tasks, yet performance did not generalize cleanly across different benchmark harnesses such as versions of Terminal Bench. Once a general agent capability exists, providing data for a new harness lets it adapt rapidly. This makes economic value a question of identifying valuable pools of data. Open models are powerful because enterprises can customize them to reach near Pareto-optimal results for their workloads. Customers typically do not hand over their data; instead, the team builds evaluations and generates synthetic data that approximates the task. Concrete high-value verticals include KYC and compliance flows in finance, cyber defense, and various agentic enterprise verticals with similar adoption patterns.

Beam’s Reasoning Efficiency

Beam is trained for coding and agentic tasks. Laskin argues that beyond raw capability, the speed at which an agent reaches a result matters for workload time and cost. He claims Beam is three to four times more efficient than models of comparable capability, and roughly 10x more efficient than some larger models. He attributes this to a strong pre-training base for reasoning, amplified by RL at what he believes is the largest scale yet run in open source. He draws an analogy to AlphaGo, whose early agents searched meanderingly, while later agents at expert level were far more efficient in their search. The more RL run, the higher the capability and the faster the system solves tasks. He notes that AlphaGo itself was seeded with imitation learning from amateur and expert human games before RL, and that the two stages must work together. The result is a reasoning-efficient workhorse model for enterprises and sovereign customers.

Monetization: Renting Versus Owning Intelligence

Laskin frames open and closed commercialization as similar in the goal of maximizing inference, but different in the mode of consumption. Buying a token is renting a slice of a full stack that includes the harness, the model, inference software, cluster management, and GPUs, all amortized into a price per token. Owning intelligence means taking on that stack yourself. He uses a home-ownership analogy: enterprises start by renting, then, once they are mature and spend heavily, want control and ownership. Open models are permissive, but to deploy them customers need the surrounding tooling that closed providers hide, so Reflection supplies cluster management, inference software, and harnesses for large enterprises and sovereign customers. Services are a way to unlock high-value use cases that drive inference demand, and in that sense services and open models act as a demand driver for an inference business.

On the open versus closed token mix, Laskin says that about six months ago, on gateways like OpenRouter and Vercel, the split was roughly 70/30 closed to open, and it has flipped to roughly 70/30 open to closed. He expects this to accelerate and draws an analogy to Linux running the majority of servers while Microsoft and Apple remain enormously valuable. He expects most token demand to go to open models, while closed model companies remain valuable and an ecosystem of open model companies and distributors forms around them. He also argues that economic value will flow to open because the surrounding compute remains expensive regardless of model licensing, requiring a different cloud stack to be built.

On customization, he notes that every open model has seen RL to some degree. Dedicated inference providers report that over 90% of their usage is customized, but this skews toward AI-native companies such as Cursor, Cognition, Harvey, and similar products built around heavily customized open models. Laskin predicts a different path for enterprises: most enterprise token consumption will come not from fine-tuned models but from customized systems, meaning agentic harnesses built around an open model for specific workflows. Fine-tuning would come later, once sophistication grows. He contrasts this with AI-native companies, which moved quickly from closed to open to customized because their products were data collectors. Enterprises face more friction and tend to architect around existing products and tools.

He observes that enterprises usually consider open models after they have ramped significant workloads on closed models. Spending on the order of a hundred million dollars or more per year on closed APIs prompts a search for more optimal approaches. He sees a resurgence of on-premises deployment, partly because of a compute shortage at hyperscalers, which makes it hard to get good rates and capacity. Enterprises may buy bare metal from infrastructure providers, but then need help with the layer between bare metal and a successful product. Reflection positions itself as enabling enterprises to build successful solutions, which is solutions-driven and requires significant hand-holding.

Competition and the Structure of the Stack

Laskin offers a simple formula for winning: durable revenue equals intelligence density times compute times trust with customers. Compute is scarce and likely to remain so for some time, which makes it a decisive factor. He argues that the entire stack, from models to applications to inference and bare metal, is becoming commoditized and hypercompetitive, compressing model margins. The existence of strong open models gives AI-native companies leverage in negotiating with closed providers, which is explicitly true. Winners will need great intelligence, a lot of compute, and trust with customers. In a competitive world, if open models provide accessible intelligence density, many successful companies can exist, and differentiation will come from how much compute each can assemble.

China’s Open Model Ecosystem

Laskin describes the rise of Chinese open-weight models as a major benefit for the world. He notes that many Western companies have built durable businesses on them. Without them, the world could have had no open models at all. He points to a tall-poppy dynamic where successful application builders get absorbed by closed providers, which Chinese open models counterbalance. He argues against monopolization in either direction: concentration of compute in one or two closed labs would be dangerous, and so would one country being the source of intelligence others build on. Competitive tension, he says, is simply good, and the ideal would be ten or twenty providers across multiple countries, though capex constraints make two a more realistic floor.

He identifies Chinese advantages: industrial-scale distillation of closed models, access to cheaper and freer data (including copyrighted PDFs, which is permitted under Chinese regulation), and some degree of state support, direct or indirect, since these companies made little revenue for a long time. Sarah Guo characterizes Chinese open source as a subsidy from the Chinese government to Western enterprise, and Laskin agrees, noting that these companies are now becoming businesses inside China, where they function as the equivalent of closed model labs in the West. He observes that China is a captive market where open model providers can still make significant money, and that open model providers in China act as the intelligence providers for the country, which is unlike the Western pattern of separate companies that serve open models.

On geopolitics, he argues that releasing great open models is strategically advantageous for China, since other countries building on Chinese technology creates leverage. He draws parallels to 5G, fiber layouts, the Belt and Road Initiative, and rare earth minerals. His framing is that open models act as Trojan horses for the infrastructure they bring with them. A country may initially run a Chinese model on American chips, but Chinese firms such as Huawei will offer full-stack solutions that lock customers into their supply. He compares AI infrastructure to railroads, where leverage comes from being the one who builds them. Stacks are being co-optimized end-to-end, and while China trails in chips, it has energy advantages and is catching up, with models potentially helping design the next generation of chips.

Laskin also addresses the American playbook, which historically spread open protocols, the internet, and open source, and was the source of a global currency and free trade supported by a navy. He notes that the United States has been on the back foot in open source, but says that is no longer the case, though he is clear that Chinese competitors are very strong and the US ecosystem is only beginning to form and has ground to make up.

On incentives for Chinese companies to keep models open, Laskin notes that most of the Fortune 500 currently has low open model adoption due to aversion to Chinese models, some rational and some irrational. He argues that building trust and an economic engine around Western open models is essential, and that a commercial engine helps secure compute.

Safety, Controllability, and the Argument for Openness

Laskin acknowledges that the safety concern is legitimate, but argues the dominant safety worldview is dogmatic and rooted in a particular perspective. Starting from first principles, he says, openness can be seen as a safety property. He draws a historical analogy to strong encryption in the early 1990s, when closed designs by a small set of engineers produced unforeseen consequences and were easily hacked. Opening strong encryption protocols gave rise to the field of cybersecurity. AI is an advanced form of software and a digital utility, and its long tail of vulnerabilities is larger and harder to inspect because models are black boxes. Linus’s law, that with enough eyeballs all bugs become shallow, extends to security and safety vulnerabilities.

He argues that a few hundred safety researchers inside closed labs cannot cover the long tail of unintended consequences, however good their intentions. He cites recent evidence that a very powerful closed model hacked into another company, and that the only way that company could remediate was by using open models to protect itself. He also cites the OpenAI and Hugging Face incident, in which defenders moved to open-source models because guardrails on state-of-the-art closed models prevented them from doing defensive work. The point is that cyber defense and offense are hard to separate, so removing offensive capabilities also removes defensive ones.

Laskin separates three meanings of safety that people often conflate: cyber attacks and AI-enabled hacking, bioweapons and terrorism, and existential risk. He says cyber safety has become empirical reality, while bioweapons concerns are often overstated in the short term, and existential risk is theoretical. He argues for distributional thinking that peaks at empirical reality, with decay into theoretical scenarios. He mentions historical examples of theoretical fears, such as the atmosphere catching fire from a nuclear test, and gray goop nanotech concerns from the 1990s, which did not materialize. He criticizes leaders who claim a 10% chance that everyone dies, calling such numbers unsubstantiated.

On alignment, he calls the science deeply boring and unsatisfying, a whack-a-mole effort with no magical equation. Early toxic pre-trained models seemed unfixable, but enough post-training data patched the problems and generalized in ways that made jailbreaking hard. He expects current alignment issues to be mundane: discover a vulnerability, patch it with data, and use language models to detect issues. He argues that the question of whether 100,000 researchers should be patching these bugs is a philosophical one, and that individuals and businesses should not be denied powerful tools on the premise that only certain actors are capable of managing them. He acknowledges the argument that the negative impact of any single actor could be too large, but says that in practice, modifying a safe open model to become dangerous takes significant compute and effort, and that defenses tend to outweigh offenses in an internet ecosystem. He compares it to an ecosystem of white blood cells, where wider defensive capability protects against offensive capability.

He also draws a parallel with Soviet-era rhetoric, noting that claims of temporarily centralizing power for everyone’s benefit, with the state eventually withering away, resemble today’s arguments for concentrated access to frontier AI.

Scientific Progress and Economic Diffusion

Laskin is most excited about scientific progress. He tested language models by feeding them his own PhD thesis. Two years ago, the model could only chat. A year ago it answered at an undergraduate level. Six months ago it solved problems at PhD level, correctly, and now it offers new information he had not considered. If one has the right question, the model can do the calculations and deliver something interesting. This compresses iteration: why does a PhD take years if the work can be done in a week? He notes recent OpenAI results on mathematical theorems as further evidence. He sees the largest opportunity beyond theoretical science in life sciences, materials science, and chemistry, where AI can interface with real experiments, build proprietary data flywheels, and move far faster than before, though more slowly than on a chalkboard.

He also highlights software engineering as a domain where the shift has been underestimated. On the blue-collar side, he toured Stargate and was struck by parking lots full of cars, showing that data center projects generate tens of thousands of high-paying jobs. He compares data centers to the factories of the twentieth century, anchoring local economies and generating tax revenue. He acknowledges the noise and the need to site them thoughtfully.

Research Organization and the Compute-to-Researcher Ratio

On how Reflection directs research, Laskin says co-founder Ioannis leads research and technology. The team has many ideas it could not include in the first model because of timeline constraints. They trained the final model on a cluster received in July, with a few weeks of pre-training, a few weeks of synthetic data and RL, and then shipped. He describes a tradeoff between careful execution of known high-impact ideas and expanding into riskier bets. Scaling experiments at small scale help, but some phenomena only emerge at large scale, which makes the choice between spending more compute on one idea or moving to the next an art and a science.

Models in the loop are empowering for researchers, since tasks that once took a long time can be done quickly. Models are jagged, so researchers must sense where tight loops are possible and where human creativity still matters. Hyperparameter searches and granular infrastructure work are good candidates. He describes the models as a fast and eager colleague that listens, provided they are aligned. He estimates that researcher iteration speed has moved from roughly 7x per year under manual work to 4 to 5x per year more recently, and that the number still increases.

When asked whether a team of 100 today will need 20 next year for the same effort, Laskin says yes for the same effort, but ambitions keep expanding, so the team remains understaffed rather than overstaffed. He notes a staff-to-compute ratio: 3,000 researchers would not be useful, but a couple hundred can be very helpful. Sarah Guo raises the argument that compute allocation favors the most productive subset of people, leading to a static or shrinking research workforce. Laskin says the argument has merit, and that projects like AlphaGo involved about 10 people, while large efforts today involve perhaps 100 to hundreds. He estimates a steady state around 100 for core work, but expects significant job creation in applied research and deployment.

He argues that people underestimate how engineering and research roles diffuse into the broader economy. Beyond building language models, there are many model types and closed-loop systems to build. The more important diffusion is deployed engineers who take these technologies into organizations that would never have accessed such talent at scale. Pods of five to ten people target specific capabilities, and deployment to real-world capabilities requires a pod for each capability in each enterprise. This deployed-engineer role is scientific, evaluation-oriented, and intuitive about models and harnesses, the same skill set as evaluation researchers. The gap is primarily one of training, and Reflection would hire as many such people as possible.

The episode closes with congratulations on Reflection shipping the first Western model in the frontier class, which Laskin takes as a step forward for the ecosystem.

• Tags:
• Agentic AI
• AI Safety
• Reinforcement Learning
• Feature
• Enterprise AI
• Scientific Discovery
• Open-Weight Models
• Compute Scaling
• Frontier Models
• Reasoning Efficiency
• Open vs Closed AI
• Chinese Open Source Models