Who will own the future of AI?

Summary
In this interview on the Sources Podcast, host Alex Heath speaks with Reflection AI co-founders Misha Laski and Ioannis Antonoglou about the rapid growth of open-weight artificial intelligence models. The founders discuss the technical and economic motivations behind Reflection AI, the strategy of scaling reinforcement learning from scratch rather than distillation, their partnership and compute deals, and their upcoming open-weight mixture-of-experts model, Beam.

What is shown

• [00:00] In-studio interview with host Alex Heath and Reflection AI co-founders Misha Laski and Ioannis Antonoglou.
• [05:15] Mid-roll sponsor advertisements for Mercury, Jira by Atlassian, and Granola.
• [08:04] Discussion and announcement of Reflection AI’s model Beam (Beam-501B-A23B).
• [29:13] Mid-roll sponsor advertisements for Mercury Books, Jira, Granola, and Framer.
• [37:46] Graphic overlays of news headlines referencing Reflection AI’s reported compute deals with Nebius ($1B+) and SpaceX ($6.3B).
• [48:54] Graphic overlays of reported enterprise partnerships, including Shinsegae Group (250MW data center in South Korea) and Dell Technologies.

Claims & numbers

• Alex Heath states that mentions of open models on corporate earnings calls increased sixfold (6x) in Q3, and open tokens represent the majority of traffic through Vercel and OpenRouter [00:10].
• Misha Laski claims that six months prior, token consumption on OpenRouter and Vercel was roughly 30% open models and 70% closed models, whereas it has inverted to approximately 70% open and 30% closed [02:54].
• Misha Laski states that Reflection AI’s debut model, Beam, is a 501B total parameter mixture-of-experts (MoE) model with 23B active parameters [08:58].
• Ioannis Antonoglou claims that through scaled reinforcement learning (RL) without distillation, Beam achieves 3x to 4x higher reasoning/token efficiency than models like GLM-5.3 or closed models like GPT-6 Luna [10:55].
• Alex Heath notes that Reflection AI has raised between $2B and $2.5B at a reported ~$25B valuation, alongside massive compute agreements with Nebius and SpaceX [37:24].
• Misha Laski states their commercial RL runs train on roughly 10,000 GB300 GPUs running for weeks, serving trillions of tokens and billions of agents internally [59:20].
• Ioannis Antonoglou states that Reflection AI plans significantly larger base models and further scaled RL runs for 2027 [57:33].

Notable quotes

• [01:25] Misha Laski: “Closed models are the equivalent in real estate to renting an apartment... The only way to own intelligence is, by definition, if it’s open.”
• [17:34] Misha Laski: “Linus’s Law, which is that with enough eyeballs, all bugs become shallow. And that’s why open-source software is actually in many ways considered safer and more trustworthy.”
• [49:55] Misha Laski: “The revenue potential of a company follows a pretty simple equation. It’s intelligence density times the amount of compute that you have, times the trust that you have.”

Assessment
This is a standard long-form podcast interview rather than a live product demonstration or technical release video. No live benchmarks, terminal sessions, or model inference outputs are directly demonstrated on screen; all technical specifications and performance comparisons are verbal claims made by the founders.

Described by gemini-3.8-flash on 2026-10-10 from the video’s audio and frames.