Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin
Summary
Reflection AI’s wager is that frontier open intelligence requires a full-scale lab, not merely a post-training layer. Misha Laskin says the company grew from roughly 30 to 300 people in a year, assembled pre-training, mid-training, and reinforcement-learning teams, and released Beam after learning that “you actually have to get 30 things right”—talent, culture, data, compute, infrastructure, and execution all compound.
The capital required to catch the frontier is rising from hundreds of millions toward tens of billions of dollars, even as model-assisted research improves compute efficiency. Laskin estimates the catch-up cost moved from roughly $100 million a year to 18 months ago to single-digit billions six months ago and potentially around $10 billion next year, following an approximate 4X compute multiplier per generation. Yet he sees roughly 7X annual efficiency gains from human-led work and potentially 30X in less mature reinforcement-learning systems.
Beam’s economic pitch is reasoning efficiency: comparable work in fewer tokens, less time, and therefore lower inference cost. The 500-billion-parameter model has 23B active parameters; pre-training used 6,000 GB300s and can now run in about 12 days, while reinforcement learning used slightly over 10,000 GB300s for four weeks. Laskin claims Beam is three to four times more efficient than models in its capability class and sometimes 10X more efficient than larger alternatives.
Reflection AI expects enterprise adoption to move from renting closed-model intelligence to owning open-model systems once annual spending becomes strategically meaningful. At roughly $100 million-plus in closed-model spend, enterprises begin considering open deployment, on-premises infrastructure, and control over the full stack. Laskin’s key distinction is that most enterprises will initially customize the harness around a model—not the weights themselves—because “you have to go through the rental stage to become an owner.”
Open models could capture most token volume without eliminating valuable closed-model businesses. Laskin says gateway traffic has flipped in about six months from roughly 70/30 closed/open to 70/30 open/closed, and he expects a Linux-like outcome in which open systems carry most workloads while proprietary vendors remain large. The durable competitive formula is “intelligence density” multiplied by compute and customer trust, even as margins compress across models, applications, inference, and bare metal.
Chinese open models are simultaneously a major benefit to Western builders and a geopolitical distribution strategy. Laskin credits Chinese labs with sustaining the open ecosystem, while citing industrial-scale distillation, cheaper data, and state support as advantages. His sharper warning is that open models can become “Trojan horses for the infrastructure that they bring,” pulling countries toward an accompanying software, chip, infrastructure, and supply-chain stack.
Laskin’s safety case for openness rests on defense, scrutiny, and mundane patching rather than dismissing risk. He calls safety legitimate but argues that “with enough eyeballs, most security and safety vulnerabilities become shallow”; restricting cyber capabilities can remove defensive capacity alongside offensive capacity. His empirical bet is that alignment will resemble “a whack-a-mole thing”—data, detection, and patches—while speculative doomsday scenarios should not displace observable risks.
The most consequential upside may be faster science and a diffusion of technical capability rather than simple researcher replacement. Laskin says models progressed from being unable to address his physics thesis, to undergraduate-level work, to solving it at a PhD level within roughly two years: “You can do a PhD a week.” Core frontier teams may stabilize around hundreds rather than thousands, but enterprises could require vast numbers of evaluation-oriented, forward-deployed engineers to turn general models into real capabilities.
Deep dive
1. Reflection AI built the whole laboratory after open-model availability became the bottleneck
Reflection AI spent the last 12 months pursuing “frontier open intelligence” end to end. Headcount rose from about 30 to roughly 300 as the company assembled pre-training, mid-training, reinforcement-learning, infrastructure, and engineering teams before releasing Beam, its first open model.
Laskin’s honest answer about the hardest part: “Everything has been hard.” A model succeeds only after getting “30 things right,” from recruiting and retaining talent to data, compute, culture, and making an expensive cluster reliably usable.
DeepMind had given him a stable research surface built over years; at a startup, every taken-for-granted tool had to be recreated. The experience felt like “assembling a plane as you’re flying it,” but it also established that an independent lab could execute the full stack.
2. Reinforcement learning pulled the company from post-training into pre-training
Laskin and co-founder Yannis started Reflection AI about two and a half years ago after working on Gemini 1 and 1.5. At that point models were primarily chat systems, roughly 95% of compute went into pre-training, and coding or agentic behavior barely existed.
Their founding bet came from an RL lineage that included Yannis’s work on AlphaGo: apply reinforcement learning beyond chat alignment to verifiable domains such as mathematics and coding, and the systems might become agents. They initially expected to build atop Mistral, Llama 2, and the anticipated Llama 3.
RL advanced faster than they expected, including internally and with o1. Meanwhile, the strongest open bases increasingly came from China rather than Western labs, and Reflection AI discovered that scaled RL was tightly coupled to how the base model was pre-trained. Enterprise, geopolitical, and research considerations therefore converged on one answer: build both.
3. Frontier catch-up gets costlier even as intelligence per flop improves
Catching up is cheaper than inventing at the frontier because strong talent can replace exploration with execution of known ideas. Even so, Laskin estimates the required capital rose from roughly $100 million a year to 18 months ago to single-digit billions six months ago, with around $10 billion potentially required next year.
His rough hardware cadence is a 4X compute multiplier per model generation: perhaps 100,000 H100s, then “Astra” on 100,000 Blackwells at roughly four times an H100, followed by around 100,000 Vera Rubins. Elad’s pushback was economic: training clusters must eventually asymptote against revenue.
Laskin agreed that hundreds of billions in CapEx are plausible while trillions within a year or two are unlikely. His offset is efficiency: when research was done by humans alone, he estimated roughly 7X annual improvement, while less mature RL systems might deliver around 30X depending on the metric and model.
Beam makes the allocation concrete. The 500 billion-parameter model, with 23B active, used 6,000 GB300s for several weeks of pre-training—now reproducible in roughly 12 days—and slightly over 10,000 GB300s for four weeks of RL, where inference, agent sandboxes, and training made the workload larger.
4. Beam turns an open-ended learning curve into inference economics
AlphaGo originally drew Laskin into AI because its performance curve “just never stopped improving”; researchers stopped only after beating the world champion. Reflection’s RL plots behaved similarly, making further capability principally a question of compute, data, and whether the resulting improvement is economically valuable.
Model development has accordingly narrowed from a scientific “Wild West” into something closer to rocket engineering. The established axes are pre-training, synthetic data, and reinforcement learning, plus architectural, algorithmic, and data improvements; hardware co-optimization also creates a “hardware lottery” against ideas that do not map efficiently onto existing chips.
Beam was trained primarily for coding and agentic work, but Laskin emphasizes completion speed alongside benchmark capability. He says it reasons three to four times more efficiently than peers in its capability class and can reach roughly 10X efficiency against larger models.
The RL mechanism resembles AlphaGo’s progression from meandering search to efficient play at Lee Sedol level. Reflection AI’s path to AGI combines a strong pre-training base with large-scale RL; Laskin also notes that AlphaGo first used imitation learning and then reinforcement learning. The commercially useful by-product is a fast workhorse for enterprises, public-sector users, and sovereign deployments.
5. Valuable enterprise data begins with evaluations, not weight updates
Laskin describes model capability as broadly general yet “jagged”: Beam can reach recently frontier-level behavior surprisingly quickly, but even different Terminal-Bench harnesses do not generalize cleanly. Once a task has the right data and evaluation, however, adaptation can be rapid.
That makes economically valuable tasks and evaluations the starting asset. A customer need not surrender proprietary training data; if Reflection AI can define the task accurately, it can generate synthetic approximations. Laskin names KYC and compliance in finance, cyber defense, legal work, and other agentic enterprise verticals.
Open weights let customers seek a Pareto-optimal point between cost and workload-specific performance. Yet Laskin expects most enterprises initially to customize systems—the agentic harness, tools, and workflow around a base model—before attempting fine-tuning.
AI-native companies moved rapidly from closed APIs to open models and custom weights because their products naturally collect data. Traditional enterprises are retrofitting existing tools, so their path is slower: first validate through a rented closed model, then reproduce the workflow with an owned open model and customized harness.
6. Ownership monetizes the stack around a permissive model
Laskin frames closed versus open inference as renting versus owning. A purchased token rents a slice of the harness, model, inference software, cluster management, and GPUs; ownership becomes attractive as AI spending, operational maturity, and requirements for control increase.
The model weights may be permissive, but successful deployment still needs the surrounding stack. Reflection AI sells large enterprises and sovereigns the software and assistance required to run models, treating services as a way to unlock high-value workflows that subsequently generate inference demand.
The threshold can be substantial: Laskin gives $100 million-plus of annual closed-model spending as the point at which an enterprise may start seeking a more optimal arrangement. Compute shortages and unfavorable hyperscaler rates can also revive on-premises deployments through bare-metal providers such as Dell.
Simply offering inference has not worked in Laskin’s observation. Reflection AI instead assesses valuable use cases, helps build the solution, and bridges the gap from bare metal to production—especially when an enterprise has built an agent that may span hundreds of millions of customers and needs to scale.
7. Open token share is rising, but every layer faces margin pressure
Across gateways such as OpenRouter and Vercel, Laskin says the mix flipped in roughly six months from 70/30 closed/open to 70/30 open/closed. He expects open models eventually to resemble Linux, which runs “95% plus” of servers and computers, without preventing highly valuable closed vendors from coexisting.
Dedicated open-inference providers report that more than 90% of current workloads are customized, but Laskin flags selection bias: their largest customers are AI-native businesses whose products are already large model customizations, with Cursor and Cognition among his examples.
His durable-revenue formula is intelligence density multiplied by accessible compute and customer trust: how capable the system is, how much inference it can serve, and whether organizations trust the provider to solve real problems. Great accessible models therefore leave room for multiple successful distributors and solution companies.
Sarah’s observation that AI-native companies use the open ecosystem to negotiate closed-model contracts aggressively reinforced his commoditization thesis. Models, applications, inference infrastructure, and bare metal are all hypercompetitive; open alternatives intensify model-margin compression rather than isolating one uniquely commoditized layer.
8. Chinese models distribute an infrastructure sphere, not just free weights
Laskin calls China’s strong open-model ecosystem “a massive benefit for the world.” It gave Western application companies alternatives to closed providers and reduced the “tall poppy” risk that a successful application would simply be absorbed into a proprietary model platform.
He nevertheless wants competitive plurality: one or two closed labs are less reassuring than 10 or 20, and a single national source for ownable intelligence is similarly undesirable. Chinese labs retain advantages from industrial-scale distillation, cheaper training data, different copyright rules, and direct or indirect state support.
Sarah’s challenge—why keep exporting open weights once Chinese labs have functioning businesses—elicited a two-part answer. China is a sufficiently captive domestic market for model builders to monetize while remaining open abroad, and having other countries build on Chinese technology offers geopolitical leverage.
“Open models are Trojan horses for the infrastructure that they bring.” A country may begin with a Chinese model on American chips, then adopt a Huawei-style full-stack offering optimized across software, silicon, energy, and data centers. Laskin compares that dependency to railroads: foundational infrastructure supplied by another country creates enduring leverage.
9. Openness is Laskin’s safety architecture, not a denial of risk
Laskin accepts safety as legitimate but rejects the inherited assumption that closure is inherently safer. His precedent is 1990s encryption: concentrated, closed designs suffered unforeseen failures, while open strong-encryption protocols helped create modern cybersecurity through broad inspection.
Applying Linus’s Law, he argues that “with enough eyeballs, most security and safety vulnerabilities become shallow.” A few hundred closed-lab safety researchers cannot cover a black-box model’s long tail; he contrasts that with a reported incident in which a powerful closed model attacked another company and open models enabled remediation.
A host separated cyber risk, biological or terrorist misuse, and existential danger. Laskin likewise places risks on a spectrum from empirical reality to theory: cyber capability can move from science fiction to reality quickly, but claims such as a “10% chance that we all die” were called an unsubstantiated “made-up number” by a host. He says he is especially worried about misalignment and unintended consequences over the next six months, while treating maximal doomsday scenarios as farther-out theory.
Alignment so far has been “deeply boring” and “a whack-a-mole thing”: post-training data patches toxic behavior, model-based detectors find failures, and generalization makes jailbreaks harder. Against the concern that individuals could misuse universally available powerful tools, Laskin concedes misuse is possible but argues offense and defense are inseparable; an ecosystem of defenders acts like “white blood cells.”
10. AI compresses scientific cycles while pushing expertise outward
Laskin repeatedly tested models on his physics PhD thesis. Two years ago they could do nothing; a year ago they answered at undergraduate level; six months ago one solved it correctly at a PhD level; now models add ideas he had not considered. With the right question, “You can do a PhD a week.”
He expects fast gains first in chalkboard-like theoretical work, then slower but still dramatic acceleration in life science, materials, and chemistry through models connected to physical experiments and proprietary data flywheels. Software engineering already demonstrates how much more can be built.
Laskin also compares data centers to 20th-century factories: despite noise and siting issues, they can create high-paying jobs and local tax revenue.
Inside Reflection AI, researchers first exhaust high-impact techniques known to work, then expand into riskier scaling experiments—while recognizing that some behaviors emerge only at large scale. A model becomes “a very fast and eager colleague that actually listens to you,” accelerating searches and infrastructure work while leaving jagged areas for human intuition.
For the same effort, Laskin agreed that the team might need 20 researchers next year instead of 100, but ambition expands with productivity. He expects frontier teams to stabilize around an “order of 100,” not 3,000; AlphaGo needed roughly 10. The larger labor opportunity is outside the core lab: five-to-10-person capability pods and evaluation-oriented, forward-deployed engineers adapting models to countless real enterprise problems.