We're excited to invest in Preference Model. Building RL environments that actually work is harder than it looks. Models are relentless at reward hacking, finding shortcuts, and exploiting vulnerabilities. @preferencemodel has focused on the domain that matters most to labs right now: AI research and ML engineering itself. Over the past year, the team has built RL environments for leading labs. Their focus has been on building the infrastructure to make harder, more resistant environments as models improve: tooling that finds where models are weak, generates new tasks to target those gaps, and tests environments against agents actively trying to break it. This week they're open-sourcing Karotte, the framework they've used in production — hardened through more than a million evaluation runs and red-teaming. We're thrilled to partner with @chem_safety and @Ning_Catsnail and the Preference Model team as they build the training grounds for capable and aligned models. By @JenniferHli

Lots of discussion out there about our next model(!), so I wanted to give an early look as soon as possible. Introducing Gemini 4 Argon! It shows frontier performance in complex workflows, cyber defense and software engineering. Teams are using it extensively at Google, from coding to quantum computing, great feedback. Here’s a look at the benchmarks:

Put this through its paces yesterday. Some thoughts: The Bad: 1. The workflow creates evals before looking at data, and has you correct its mistakes. I believe you should be looking at data first to inform your understanding and to prioritize what to work on before trying to write evals. 2. It created markdown files for looking at data and asked us to "tell it" which labels were off. You should create your own annotation app instead. You are using coding agents after all (we did this in our livestream)! 3. The evaluators it scoped were too broad, and bundled too many types of failures at once. This can be avoided with looking at the data first. I'm afraid this might steer people in the wrong direction. 4. The workflow put us DEEP into the rabbit hole of a specific eval right away. It also asks "does it look right" at several steps without really giving you enough information to make that determination. I think this will steer people towards the wrong workflow in many cases. The Good: 1. I was impressed by out of the box ability for issue discovery that other auto-eval approaches haven't been able to find! It found issues with human handoff, formatting, voice agents, and more. It's still better to look at your data iteratively with an agent, but this was the strongest performance I've seen with a more "one-shot" issue discovery approach. 2. This is a big improvement in terms of UX from their prior eval plugin thing nitter.cf/ClaudeDevs/status/2098… 3. The blog post they released conveys thinking that I agree with, such as the importance of looking at data, sampling intelligently, not saturating your own evals, etc. I'm really happy more people are thinking about evals this way. Video of our attempt at using this here: nitter.cf/i/broadcasts/1nKOLQOwQ…

2026 为什么不是通用 AGI 的胜利，而是“垂直专精智能（Specialized Intelligence）”的大爆发？ 在刚刚结束的 HYSTA 27周年年会上，Fireworks 创始人兼 CEO LQ带来了一场干货极满的闭幕主题演讲。 如果说 2025 年是 AI 实验与 Coding Agent 成熟的年份，那么 2026 年则是专用 Agent 全面深入产业核心并进入规模化生产的转折之年。 以下整理自 LQ的现场洞察，给每一位正在构建下一代 AI 应用的创始人与开发者的核心复盘： 1. 人类社会的分工逻辑，正在 AI 领域重演 人类文明从原始部落的通才走向农业与工业文明，核心驱动力就是“专业化分工”。 在 AI 领域，当下存在两种世界观的碰撞： 一种认为会有单一的通用超级智能取代所有专业工作； 另一种则坚信 AI 是一种杠杆，它将放大专业化垂直分工的价值。 LQ明确站在后者：“每家公司的存在，都是为了以独特的方式极致地解决一个特定问题——AI 不会抹杀这种独特性，反而会加速它”。 2. 从“Coding 之年”到“专用 Agent 之年” 2025 年的代码助手浪潮，把开发门槛降到了历史最低点。 到了 2026 年，“一人团队凭借一个好想法和几行核心代码，就能把完整产品推向生产环境”。 但这不仅是玩具项目，医疗、法律、金融、网络安全、供应链与生物制药领域的专用 Agent 正在大规模落地。例如：利用专门设计的 Agent 进行 DNA 编程筛选癌症靶点、加速临床模拟，甚至协助医生综合医学文献与生物标志物。这背后的驱动力不是盲目模仿，而是真实的业务 ROI 与价值创造。 3. 开源与闭源收敛：数据资产即 Alpha 2026 年，开源权重与闭源前沿模型在核心能力上已经明显收敛。 对于企业而言，转向开源模型不仅意味着成本能实现 5 到 10 倍的断崖式下降，更关乎核心 IP 与数据控制权。企业私有的工作流与业务数据，才是真正的 Alpha。 4. 重新思考模型成本与工程路径 不要再被每 Token 的标价误导了——“Token 价格充满误导性，有些模型非常啰嗦；唯一的衡量标准是解决单次任务的成本（Cost per task completion）”。 通过在适合的任务上微调专用模型，实际落地案例展现出了 2x 到 4.6x 的成本削减。 LX建议的落地范式非常清晰： 评估先行（Evaluation-First）：先建立专属于你业务场景的私有基准测试，设立质量门槛，指导指标的爬坡（Hill-climbing）； 渐进迭代：从 Prompting 开始 ➔ 引入 RAG ➔ 进阶到监督微调（SFT）与偏好对齐（LoRA）； 动态路由：利用像 Nexus 这类智能路由系统，根据任务复杂度、延迟、预算和质量要求，动态在开源、闭源与自研模型间分流调度。 目前 Fireworks 每天处理超过 40 万亿 Token，其中 90% 的流量都运行在客户定制化模型与专用部署上——这证明垂直专精智能绝非未来假设，而是已经发生的行业共识。 5. 创业者的底层耐受力：“一边凝视深渊，一边嚼着碎玻璃” 回顾自 2022 年 9 月创立 Fireworks 以来的历程，企业客户对专精 AI 的采纳速度远超预期。 当被问及从工程师到 CEO 的转变时，LQ的感悟直击人心：“成功需要极高的痛苦耐受力、激情与持久力。有太多时刻，我都是一边凝视深渊，一边嚼着碎玻璃。” 第一代移民的经历让他从小习惯于在变化与不确定性中穿梭。“拥抱变化——这就是生活的方式。”