AI Builders Digest — 2026-10-11
X / TWITTER
Thibault Sottiaux, Codex & ChatGPT at OpenAI
Sottiaux says a ChatGPT subscription now doubles as a Devin subscription, pointing to a tighter bridge between general AI subscriptions and autonomous software engineering. He also flagged a major upgrade to Dots, although the post itself did not provide technical details.
Sottiaux 表示，ChatGPT 订阅现在也相当于 Devin 订阅，说明通用 AI 订阅与自主软件工程工具正在加速融合。他还提到 Dots 获得大幅升级，但原帖没有披露具体技术细节。
Sources: ChatGPT + Devin · Dots upgrade
Nan Yu, Product at OpenAI Codex
Yu describes the next turn in coding interfaces: developers moved from writing code manually, to tab-completing code, to prompting agents, and are now tab-completing the prompts sent to agents. The implication is that intent specification itself is becoming an assisted, iterative interface.
Yu 描述了编程界面的下一次演进：开发者从手写代码，到补全代码，再到 prompt agents，如今连发给 agents 的 prompt 也开始被自动补全。这意味着「表达意图」本身正在成为一种由 AI 辅助、持续迭代的交互界面。
Source: https://x.com/thenanyu/status/2108671762984731037
Madhu Guru, Senior Director of AI at Meta
Guru argues that open-weight models did not begin the decline in the price of intelligence; they accelerate a trend already driven by distillation, infrastructure efficiency, and provider competition. His recurring pattern is that each new mid-sized model reaches the intelligence of the previous generation's large model at a lower inference cost.
Guru 认为，单位智能价格下降并非始于 open-weight models；它们只是加速了由 distillation、基础设施效率提升和供应商竞争共同推动的趋势。一个反复出现的规律是：新一代中型模型能以更低推理成本达到上一代大型模型的智能水平。
Source: https://x.com/realmadhuguru/status/2108618266776387886
Thariq, Claude Code at Anthropic
Thariq revisited a side project originally built with Opus 4 and the Agent SDK. It required an always-running process and was unreliable; a single prompt to Opus 5.5 ported it to Claude Managed Agents and materially improved reliability. The project is now receiving outside pull requests, a small but concrete example of agent infrastructure turning prototypes into maintainable products.
Thariq 重新改造了一个最初用 Opus 4 和 Agent SDK 开发的副项目。旧版本需要常驻进程且可靠性不佳；只用一个 prompt，Opus 5.5 就将其迁移到 Claude Managed Agents，并显著提高了稳定性。项目现在已经开始接收外部 PR，是 agent 基础设施把原型转化为可维护产品的一个具体案例。
Sources: Managed Agents port · Community PRs
Guillermo Rauch, CEO of Vercel
Rauch shared unusually concrete evidence of an agent-native web: 58.18% of Vercel network traffic is bot-originated, versus 32% in January 2024; more than 60% of deployments are now agentic, up from roughly 3% in January 2026; and agents account for as much as 83% of pageviews on Vercel documentation. He also says agents already purchase infrastructure through Vercel's CLI, with domain purchasing now extending the path from idea to online business.
Rauch 给出了少见的 agent-native web 实证数据：Vercel 全网流量中 58.18% 来自 bots，而 2024 年 1 月仅为 32%；超过 60% 的部署现在由 agents 驱动，2026 年 1 月约为 3%；Vercel 文档站最高 83% 的页面浏览来自 agents。他还表示，agents 已经会通过 Vercel CLI 购买基础设施，现在连域名购买也被纳入从想法到线上业务的完整链路。
Sources: Agent traffic statistics · Agents buying infrastructure and domains
Aaron Levie, CEO of Box
Levie predicts that background agents, agent spawning, and swarms will consume 1,000 times more tokens than one-at-a-time human prompting. His key claim is that chat-paced AI will soon look obsolete: most token demand will come from continuous work embedded in workflows, requiring far more compute and infrastructure than today's adoption curve suggests.
Levie 预测，后台 agents、agent 自我派生和 agent swarms 消耗的 tokens 将比人类逐次 prompt 高出 1,000 倍。他的核心判断是：以聊天节奏运行的 AI 很快会显得过时；未来大多数 token 需求将来自嵌入工作流的持续任务，因此所需算力与基础设施将远超当前采用曲线所显示的规模。
Source: https://x.com/levie/status/2108750943680630893
Peter Yang, AI educator and product thinker
Yang argues that the market has overproduced personal assistants for email while underinvesting in AI companies that manage and cure disease. On a more tactical level, he recommends giving an AI chief of staff its own identity and email name, because copying “yourself” into scheduling threads creates an awkward human-agent interface.
Yang 认为，市场已经过度投入邮件型个人助理，却对利用 AI 管理和治疗疾病的创业公司投入不足。在更战术的层面，他建议为 AI chief of staff 设置独立身份和邮箱名称，因为在排期邮件中抄送「自己」会造成别扭的人机交互。
Sources: AI startup priorities · AI chief-of-staff identity
Nikunj Kothari, Partner at FPV Ventures
Kothari warns founders against paying for influencer-style “collaborations” targeted at VCs. He sees this as a negative fundraising signal: attempts to manufacture attention can suggest poor judgment, and reputation travels quickly through investor networks.
Kothari 提醒创始人，不要购买专门面向 VC 的 influencer 式「合作推广」。在他看来，这会形成负面的融资信号：刻意制造关注可能暴露判断力问题，而且这种名声会在投资人网络中迅速传播。
Source: https://x.com/nikunj/status/2108616889605951834
Matt Turck, Partner at FirstMark
Turck highlights Synthesia's arrival at one-prompt, professional-grade company and product videos in a user's own style. His broader point is that a long-standing generative-video vision is crossing from gradual progress into broadly available product capability.
Turck 关注到 Synthesia 已能通过一个 prompt，按用户自身风格生成专业级公司或产品视频。他的判断是，一个长期存在的生成式视频愿景，正在从渐进改良跨越到大众可用的产品能力。
Source: https://x.com/mattturck/status/2108561284635480443
Dan Shipper, CEO of Every
Shipper points to Slack as an emerging operating surface for agents, where they can participate inside existing team communication rather than requiring users to switch to a separate AI interface.
Shipper 将 Slack 视为正在形成的 agent 操作界面：agents 可以直接进入现有团队协作流程，而无需用户切换到独立的 AI 界面。
Source: https://x.com/danshipper/status/2108588611000275009
PODCASTS
No Priors: Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin
The Takeaway: Reflection AI is betting that open intelligence wins not merely by matching closed models, but by combining efficient reasoning, customer ownership, and a full deployment stack.
Reflection AI co-founder and CEO Misha Laskin, formerly a reinforcement learning researcher at Google DeepMind, says the lab grew from roughly 30 to 300 people in a year to build Beam end to end. Beam is a 500-billion-parameter mixture model with 23 billion active parameters, trained for coding and agentic tasks. Laskin says it reaches comparable outcomes three to four times more efficiently than models in its capability class, with even larger gains against bigger models. Reflection used about 6,000 GB300 GPUs for pre-training and more than 10,000 for four weeks of reinforcement learning.
His commercial thesis is ownership. Enterprises begin by “renting intelligence” through closed-model tokens, then shift toward owning models and infrastructure once workloads become strategic and expensive. Reflection plans to monetize the surrounding stack: inference software, cluster management, agent harnesses, and deployment services. Laskin expects customized systems to matter before customized models for most enterprises, because adapting the harness around an open model is easier than immediately fine-tuning it.
The contrarian safety argument is equally important. Quoting Linus's law, Laskin says, “with enough eyeballs, most security and safety vulnerabilities become shallow as well.” He believes open scrutiny can expose the long tail of failures that a few hundred researchers inside closed labs cannot anticipate. His forecast is that most tokens will eventually run through open models, much as most servers run Linux, while valuable closed-model companies continue to exist.
核心结论： Reflection AI 的赌注不是让 open intelligence 仅仅追平闭源模型，而是通过更高的推理效率、客户对智能资产的所有权，以及完整部署技术栈取得优势。
Reflection AI 联合创始人兼 CEO Misha Laskin 曾在 Google DeepMind 从事 reinforcement learning 研究。他表示，公司一年内从约 30 人扩张到 300 人，以端到端方式开发 Beam。Beam 是一个总参数 500B、激活参数 23B 的混合模型，专攻 coding 和 agentic tasks。Laskin 称，其完成同等任务的效率是同级模型的 3 至 4 倍，与更大模型相比优势甚至更高。Reflection 约用 6,000 块 GB300 完成 pre-training，并用超过 10,000 块 GB300 进行了四周 reinforcement learning。
他的商业逻辑是「所有权」。企业最初通过闭源模型的 tokens「租用智能」，当工作负载变得关键且昂贵后，就会转向拥有模型和基础设施。Reflection 计划从外围技术栈变现，包括 inference software、cluster management、agent harnesses 和部署服务。Laskin 判断，对多数企业来说，customized systems 会先于 customized models 普及，因为围绕 open model 调整 harness，比立即进行 fine-tuning 更容易。
其安全观点同样具有反共识色彩。借用 Linus's law，Laskin 表示：「当审视者足够多时，大多数安全漏洞也会变得浅显。」他认为，开放审查更有机会发现闭源实验室数百名研究人员无法预见的长尾风险。他预测，未来大多数 tokens 会运行在 open models 上，就像多数服务器运行 Linux；与此同时，高价值的闭源模型公司仍会继续存在。
Source: https://www.youtube.com/@NoPriorsPodcast
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders