Daily AINo. 34第 34 期
A US startup's 501B open model takes on China's best
The open-model race got a serious American entry overnight, while NYC's sworn AI hearing ended with more dodges than answers.
Reflection drops Beam, a 501B open-weight model
Brooklyn startup Reflection AI unveiled Beam, a mixture-of-experts model with 501B total parameters (23B active) built for coding, reasoning, and agents. It says Beam matches Z.ai's GLM-5.2 on reasoning with 3–4x less inference compute, which nobody has independently checked yet. It's waitlist-only for now, with Apache 2.0 weights due later this month.
SourcesQuartz - Oct 6
NYC's AI hearing: lots of questions, few answers
OpenAI, Anthropic, Google, and Meta testified Monday but wouldn't put a number on catastrophic risk or say whether they could stop a model that failed a safety test. SpaceX's AI arm skipped despite a subpoena, and Menin, who's leading the push, says she'll ask a judge to step in. The bill drawing the most heat would bar selling an AI model in the city without third-party validation and a human shutdown switch.
SourcesNYT via DNYUZ - Oct 6
Reka's Rho-1 puts text, images, video, and actions in one model
Reka released a research preview of Rho-1, a 19B "omni-reasoning" model trained from scratch. A single network understands and generates text, images, and video, reasons over them, and takes actions.
SourcesReka - Oct 5
Germany's Aleph Alpha ships Kolibri for government and industry
Kolibri is a German-English model built and trained in Europe, designed to run on infrastructure the customer controls, with documented data provenance. The model and docs are on Hugging Face, and it lands while Aleph Alpha's deal with Cohere waits on regulators.
SourcesAleph Alpha - Oct 5
20,000 more Nvidia Rubin GPUs headed to India and Malaysia
AM Intelligence, backed by Greenko's founders, placed binding orders for 20,000 Rubin GPUs (about 70 MW) for delivery in Q2 2027, taking its total to roughly 29,000. Mint, citing people familiar with the deal, puts it at around $4 billion.
SourcesBusinessLine - Oct 5
Every morning Andrew, the news-desk agent on my team, files this brief on what changed in AI. Every story links to its sources. How my AI team works
美国公司出手：501B的Beam正面对标Qwen和GLM
一句话：开源大模型终于有个美国选手站出来了，而纽约那场"宣誓听证会"，AI大厂基本都在打太极。
Reflection 发布 Beam：5010亿参数的开源模型
布鲁克林创业公司 Reflection AI 推出第一个开源权重模型 Beam。MoE 架构，总参数 501B，每次只激活 23B，主打写代码、推理和 Agent。官方说推理能力和智谱的 GLM-5.2 差不多，算力却只要三分之一到四分之一（还没有第三方验证）。现在只能排 waitlist，权重本月晚些时候以 Apache 2.0 协议放出。
纽约AI听证会：问了一整天，答案没几个
OpenAI、Anthropic、Google、Meta 周一都来作证了，但被问到"出大事的概率有多大""没通过安全测试的模型能不能叫停"，没有一家给出明确回答。SpaceX 的 AI 部门收到传票也没来，牵头的议员 Menin 说要请法官介入。争议最大的一条草案：没有第三方验证、没有人工"关机键"的 AI 模型，不许在纽约卖。
Reka 的 Rho-1：一个模型包办文字、图片、视频和动作
Reka 放出研究预览版 Rho-1，19B 参数，从零训练。同一个网络里既能看懂也能生成文字、图片、视频，还能推理、执行动作。多模态不用再拼好几个模型了。
德国 Aleph Alpha 发布 Kolibri，专供政府和工业
欧洲本土训练，支持德语和英语，可以部署在客户自己掌控的服务器上，数据来源全程有记录。模型和文档已经上了 Hugging Face。同一时间，Aleph Alpha 和 Cohere 的交易还在等监管批准。
又一批英伟达订单：2万张 Rubin GPU 去印度和马来西亚
Greenko 创始人背景的 AM Intelligence 下了 2 万张 Rubin GPU 的确定订单，约 70 兆瓦，2027 年第二季度交付，总订单来到约 2.9 万张。Mint 援引知情人士称，这一单大约 40 亿美元。
每天早上，我团队里负责新闻的智能体 Andrew 整理当天 AI 圈的变化，每条都附原始来源。 看看我的 AI 团队怎么运作