AI Builders Digest — 2026-09-20

X / TWITTER

OpenAI's Thibault Sottiaux

Thibault Sottiaux says OpenAI is preparing a densely packed keynote with Sam Altman and Romain Huet, with some announcements arriving as early as next week and a broader set of releases converging over the coming months.

Thibault Sottiaux 表示，OpenAI 正在与 Sam Altman、Romain Huet 准备一场信息密度很高的 keynote。部分内容最快下周发布，更多新进展将在未来数月逐步汇合。

Source / 原文：https://x.com/thsottiaux/status/2101157729037586694

AI educator Peter Yang

Peter Yang calls Meta's Muse the best personal agent he has tried. He says it saved him more than $800 a year across cable and phone bills, including negotiating $288 in annual Comcast savings by calling customer support, while also handling use cases such as personalized news and habit tracking. His broader signal: consumer agents become compelling when they complete economically valuable actions, not merely answer questions.

AI 教育者 Peter Yang 称 Meta 的 Muse 是他迄今体验过最好的个人 agent。Muse 通过拨打客服电话协商账单，仅 Comcast 一项便节省每年 288 美元，全年手机和有线电视费用合计节省超过 800 美元，同时还能生成个性化新闻和追踪习惯。更重要的信号是：消费级 agent 真正的价值来自完成有经济回报的行动，而不只是回答问题。

Sources / 原文：
- https://x.com/petergyang/status/2101033599319613533
- https://x.com/petergyang/status/2101083891507593576

Claude Code's Thariq

Claude Code now supports AGENTS.md starting in version 2.1.277. When a directory has no CLAUDE.md, Claude Code will fall back to AGENTS.md; the behavior can be toggled in /config. Thariq explains that the feature is implemented as a built-in Claude Code mod, previewing a system that will let users customize the coding-agent harness and create their own project-instruction behaviors.

Claude Code 从 2.1.277 版本开始支持 AGENTS.md。当目录中没有 CLAUDE.md 时，Claude Code 会读取 AGENTS.md，并可在 /config 中开关。Thariq 还透露，这项能力基于内置 Claude Code mod，预示用户未来可以自定义 coding-agent harness，构建自己的项目指令机制。

Sources / 原文：
- https://x.com/trq212/status/2101009392611278961
- https://x.com/trq212/status/2101009393731223817
- https://x.com/trq212/status/2101009395052343462

Vercel CEO Guillermo Rauch

Guillermo Rauch reports that open models reached a possible record 78.4% of token volume on Vercel AI Gateway versus 21.6% for closed models. Spend tells a different but still notable story: Moonshot AI, DeepSeek, and Z.ai together surpassed OpenAI's second-place inference spend across providers. He also highlights Jev's rapid adoption as evidence that developers are aggressively seeking cheaper, faster AI so they can embed it in more workflows.

Vercel CEO Guillermo Rauch 表示，开放模型在 Vercel AI Gateway 的 token 用量占比达到可能创纪录的 78.4%，闭源模型为 21.6%。按推理支出统计则呈现不同格局，但 Moonshot AI、DeepSeek 与 Z.ai 的合计支出已超过排名第二的 OpenAI。他同时将 Jev 的快速普及视为一个明确信号：开发者正在主动寻找更便宜、更快的 AI，以便把它嵌入更多工作流。

Sources / 原文：
- https://x.com/rauchg/status/2101186741042663579
- https://x.com/rauchg/status/2101079472732848510
- https://x.com/rauchg/status/2101116978677285241

Box CEO Aaron Levie

Aaron Levie demonstrates Jev as a near-instant, near-zero-cost decision layer for enterprise agents. In the Box workflow, it reads an incident report, judges customer impact and severity, routes the file to the right folder, and writes structured metadata. He sees the same pattern applying to insurance claims, contracts, loans, security reviews, and customer-log analysis.

Box CEO Aaron Levie 展示了 Jev 如何成为企业 agent 的近实时、近零成本决策层：读取事故报告，判断是否影响客户及严重程度，将文件路由到相应目录，并写入结构化 metadata。他认为，这一模式同样适用于保险理赔、合同管理、贷款处理、安全审查和客户日志分析。

Source / 原文：https://x.com/levie/status/2101007708044574906

Builder Zara Zhang

Zara Zhang offers a compact rule for improving AI-era creative work: low-quality output often begins with low-quality inputs. Her advice is to fix the information diet before trying to fix the production process.

Builder Zara Zhang 给出了一条适用于 AI 时代创作的简洁原则：低质量输出往往源自低质量输入。与其先优化生产流程，不如先改善自己的信息摄入。

Source / 原文：https://x.com/zarazhangrui/status/2101123389528457596

FPV Ventures partner Nikunj Kothari

Nikunj Kothari built Jevable, a filterable gallery of Jev demos, and points to a concrete cost-performance example: scoring 3,000 children's snacks against multiple criteria in 28 seconds for $0.11. The emerging product opportunity is high-volume judgment and classification that was previously too slow or expensive to automate.

FPV Ventures 合伙人 Nikunj Kothari 构建了 Jevable，一个可筛选的 Jev demo 展示站。他还分享了一个具体的成本性能案例：用 28 秒、0.11 美元对 3,000 种儿童零食进行多维度评分。由此浮现的产品机会，是将过去因速度或成本而无法自动化的大规模判断与分类任务交给 AI。

Sources / 原文：
- https://x.com/nikunj/status/2101077053567332618
- https://x.com/nikunj/status/2101006585481073093

OpenClaw's Peter Steinberger

Peter Steinberger describes roboclaw as a live team-memory layer: it runs the team server, participates on Discord, talks through GPT Live, and tracks multiple active and historical sessions so colleagues can query context during meetings. He also recommends letting OpenClaw reorganize sessions from the home sidebar and notes that CUA can operate across these workflows more efficiently than screenshot-only interaction.

OpenClaw 的 Peter Steinberger 将 roboclaw 描述为团队的实时记忆层：它运行团队服务器、驻留 Discord、通过 GPT Live 对话，并追踪多个当前及历史 session，让团队能在会议中随时查询上下文。他还建议直接让 OpenClaw 从主页侧边栏整理 session，并指出 CUA 能比纯截图交互更高效地处理这些工作流。

Sources / 原文：
- https://x.com/steipete/status/2101141707375227372
- https://x.com/steipete/status/2101139037801283997
- https://x.com/steipete/status/2101115690719809873

Every CEO Dan Shipper

Dan Shipper pushes back against dismissing ambitious AI demos as fake, defending Jack Cheng's work as both real and a credible glimpse of a future interface. His point is less about spectacle than evaluation discipline: extraordinary demos should be verified, not reflexively believed or rejected.

Every CEO Dan Shipper 反对把激进的 AI demo 轻率斥为造假，并为 Jack Cheng 的作品背书，称其既真实，也展现了未来交互方式的一种可信雏形。核心不在炫技，而在评估纪律：面对超出预期的 demo，应当验证，而不是本能地相信或否定。

Source / 原文：https://x.com/danshipper/status/2101155521818476693

PODCASTS

No Priors: Why Diffusion Will Win AI Inference with Inception Co-Founder and CEO Stefano Ermon

The Takeaway: Inception is betting that diffusion language models will win a meaningful share of AI inference because parallel generation maps better to GPUs, lowers latency, and opens control mechanisms that autoregressive models cannot easily match.

Stanford professor and Inception co-founder Stefano Ermon helped originate score-based diffusion models before taking the architecture into text and code. Inception's 2024 research matched autoregressive quality at GPT-2 scale while generating text about 10 times faster, motivating the commercial Mercury models. Ermon argues that autoregressive inference inherits the sequential bottleneck of old RNN-style computation: each token must wait for the previous one, leaving GPUs memory-bound. Diffusion instead processes many tokens together, making inference resemble the highly parallel workload GPUs already handle well during training.

The wedge today is latency-sensitive work such as voice agents, high-volume classification, and RL rollout generation. Ermon estimates that 20% to 30% of current workloads may already be addressable where latency budgets matter most. The deeper upside is controllability: because diffusion refines an entire output from coarse to fine, an external reward or constraint can steer it throughout generation rather than only score the finished result. The cost is a new stack. Inception has had to build its own serving engine, kernels, SFT, RLHF, and RL infrastructure, creating both a moat and an adoption burden. His most memorable strategic claim is: “The bitter lesson is that the more parallel solution is the one that is eventually going to win.”

核心结论： Inception 押注 diffusion language model 将占据相当比例的 AI inference 市场，因为并行生成更适配 GPU，能够降低延迟，并提供 autoregressive model 难以实现的控制方式。

Stanford 教授、Inception 联合创始人 Stefano Ermon 是 score-based diffusion model 的早期开创者之一，之后将这一架构推进到文本与代码领域。Inception 在 2024 年的研究中，于 GPT-2 规模实现了与 autoregressive model 相当的质量，同时文本生成速度快约 10 倍，并由此发展出商业化 Mercury 模型。Ermon 认为，autoregressive inference 延续了类似早期 RNN 的串行瓶颈：每个 token 都必须等待前一个 token，GPU 因而受到内存带宽限制。Diffusion 则能同时处理多个 token，使 inference 更接近 GPU 在训练阶段擅长的高度并行工作负载。

当前切入口是语音 agent、大规模分类、RL rollout 生成等延迟敏感任务。Ermon 估计，在延迟预算决定模型选择的场景中，现有工作负载已有 20% 至 30% 可被这类模型覆盖。更深层的潜力是可控性：diffusion 从粗到细地改写整个输出，因此外部 reward 或约束可以贯穿生成过程持续引导，而不是等成品生成后再评分。代价则是必须重建技术栈。Inception 已自行开发 serving engine、kernel、SFT、RLHF 与 RL 基础设施，这既形成护城河，也增加采用门槛。他最值得记住的战略判断是：“The bitter lesson is that the more parallel solution is the one that is eventually going to win.”

Source / 原始节目：https://www.youtube.com/@NoPriorsPodcast

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders