CharlieKey takeaways from your favorite podcasts
← 9 October 2026No Priors

Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin

Misha Laskin argues that open frontier models are essential for enterprise ownership, competition and safety, and describes how Reflection built Beam through large-scale pre-training and reinforcement learning.

Watch on YouTube ↗

Charlie's summary

  1. Reinforcement learning requires pre-training, so Reflection moved from an RL bet to building open models end to end

    Misha and his co-founder started as RL researchers who planned to build on an open base model. About a year in, the good open models were coming from China, and they found that “you actually do need to pre-train your model in order to make reinforcement learning work very well at scale.” The two are tightly coupled, so they decided to build models end to end.

    ▶ 4:05
  2. Catching the frontier costs more every year, but compute efficiency gains keep the spend from running away

    Misha estimates catching up cost hundreds of millions 18 months ago, billions six months ago, and may reach tens of billions next year, roughly a 4x compute multiplier per model generation. He expects an asymptote, because models now help build themselves. Pre-training efficiency gains run about 7x a year with humans alone, and RL gains may be around 30x. Beam used 6,000 GB300s for pre-training and over 10,000 for RL.

    ▶ 8:44
  3. RL systems that never stop improving turn model building into an economics decision

    Misha got into AI after seeing a plot in AlphaGo that never stopped improving. He says Reflection's RL system shows the same curve: it keeps going up, limited only by compute. The question has shifted from whether you can train these systems to where the economically valuable data is and how much money to put in. He says this “was not true” before, and that the work now feels more like engineering than open-ended science.

    ▶ 13:18
  4. Beam is three to four times more reasoning-efficient than same-class models because of RL at unprecedented open scale

    Beam was trained for coding and agentic tasks. Misha says it is three to four times more efficient than models of the same capability class, and about 10x against larger models, so tasks finish faster and cost customers less. He credits a strong pre-training base amplified by what he believes is the largest open-source RL run, over 10,000 GB300s for weeks. He compares it to early AlphaGo agents, which meandered, versus the later, much sharper ones.

    ▶ 19:30

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.