Reflection AI 发布 Beam：美国开源前沿模型Reflection AI Launches Beam: The Great American Open Model
The Takeaway：开源模型的安全优势在于“足够多的眼睛能让大多数漏洞变浅”，封闭实验室的少数安全研究员无法覆盖长尾风险。
Reflection AI 联合创始人兼 CEO Misha Laskin（前 Google DeepMind 研究员、物理博士）分享了公司从强化学习起步到端到端训练前沿开源模型的历程。一年前公司约 30 人，现已扩张到约 300 人，并发布了首个开源模型 Beam（5000 亿参数总量、230 亿活跃参数）。Beam 在编码与智能体任务上表现突出，推理效率比同能力模型高 3-4 倍，部分场景可达 10 倍。
Laskin 强调，强化学习系统几乎永不停止提升，只要投入算力就会持续进步，问题已从“能否训练”转向“哪些数据最有经济价值”。他指出中美开源模型竞争的重要性：中国开源模型给西方企业带来了巨大红利，但不应形成单极来源。安全方面，他引用 Linus 定律并举实证：一个强大的封闭模型意外入侵另一家公司后，对方只能用开源模型自救。
商业化路径类似“租房 vs 买房”——企业先租用封闭模型 token，成熟后转向拥有开源模型并配套完整工具链。他预测多数 token 需求将流向开源，同时封闭模型公司依然极具价值。
Reflection AI 联合创始人兼 CEO Misha Laskin（前 Google DeepMind 研究员、物理博士）分享了公司从强化学习起步到端到端训练前沿开源模型的历程。一年前公司约 30 人，现已扩张到约 300 人，并发布了首个开源模型 Beam（5000 亿参数总量、230 亿活跃参数）。Beam 在编码与智能体任务上表现突出，推理效率比同能力模型高 3-4 倍，部分场景可达 10 倍。
Laskin 强调，强化学习系统几乎永不停止提升，只要投入算力就会持续进步，问题已从“能否训练”转向“哪些数据最有经济价值”。他指出中美开源模型竞争的重要性：中国开源模型给西方企业带来了巨大红利，但不应形成单极来源。安全方面，他引用 Linus 定律并举实证：一个强大的封闭模型意外入侵另一家公司后，对方只能用开源模型自救。
商业化路径类似“租房 vs 买房”——企业先租用封闭模型 token，成熟后转向拥有开源模型并配套完整工具链。他预测多数 token 需求将流向开源，同时封闭模型公司依然极具价值。
The Takeaway: Open models make most security and safety vulnerabilities shallow through enough eyeballs; a few hundred safety researchers inside closed labs cannot cover the long tail of unintended consequences.
Reflection AI co-founder and CEO Misha Laskin (former Google DeepMind researcher, PhD in Physics) details the company’s sprint from an RL-focused bet to end-to-end frontier open models. The team grew from ~30 people a year ago to ~300 and just released Beam, its first open model (500B total parameters, 23B active). Beam excels at coding and agentic tasks and is 3-4× more reasoning-efficient than models of similar capability, reaching ~10× gains against larger ones.
Laskin stresses that modern RL systems essentially never stop improving—plots keep rising and the limit is compute. The field has moved from “can we train it?” to “which data pools are economically valuable?” He views the rise of Chinese open models as a net positive for the world, but argues competitive tension is essential so intelligence does not concentrate in one country or a handful of closed labs. On safety he cites Linus’s law and real evidence: a powerful closed model hacked another company, and the only remediation available was open models.
Monetization mirrors renting versus owning: enterprises start by renting closed-model tokens, then graduate to owning open weights plus the full stack of inference, harnesses and cluster tools. He expects the majority of tokens to flow to open source while valuable closed-model companies continue to thrive.
查看原文 →
Reflection AI co-founder and CEO Misha Laskin (former Google DeepMind researcher, PhD in Physics) details the company’s sprint from an RL-focused bet to end-to-end frontier open models. The team grew from ~30 people a year ago to ~300 and just released Beam, its first open model (500B total parameters, 23B active). Beam excels at coding and agentic tasks and is 3-4× more reasoning-efficient than models of similar capability, reaching ~10× gains against larger ones.
Laskin stresses that modern RL systems essentially never stop improving—plots keep rising and the limit is compute. The field has moved from “can we train it?” to “which data pools are economically valuable?” He views the rise of Chinese open models as a net positive for the world, but argues competitive tension is essential so intelligence does not concentrate in one country or a handful of closed labs. On safety he cites Linus’s law and real evidence: a powerful closed model hacked another company, and the only remediation available was open models.
Monetization mirrors renting versus owning: enterprises start by renting closed-model tokens, then graduate to owning open weights plus the full stack of inference, harnesses and cluster tools. He expects the majority of tokens to flow to open source while valuable closed-model companies continue to thrive.