AI Builders Digest — 2026-09-28

2026-09-28

AI Builders Digest — 2026-09-28

X / TWITTER

Peter Yang — AI educator and interviewer

Peter Yang is testing Google’s new audio APIs by building a conversational Japanese tutor with 10 lessons and 10 phrases per lesson. It is a compact example of an AI-native learning product: narrow scope, voice-first interaction, and a curriculum small enough to ship and iterate quickly. He also reports that Claude usage limits have recently felt far less restrictive, though this is an individual observation rather than an official policy announcement.

Peter Yang 正在用 Google 新推出的 audio API 制作一款日语会话导师，包含 10 节课、每节 10 个短语。这是一个典型的 AI-native 学习产品：范围聚焦、voice-first，并把课程压缩到足以快速发布和迭代的规模。他还观察到 Claude 的使用限制最近宽松了很多，但这只是个人体验，并非官方政策公告。

Thariq — Claude Code at Anthropic

Thariq looks back at one of the earliest Claude Code video-making workflows from a year ago. Those results required long, painstaking iteration and repeated correction of visual details; his point is how quickly the tool has moved from laborious co-production toward much more capable generation. The post is a useful reminder that progress in coding agents is visible not only in benchmarks, but also in how many feedback loops a creator must spend to reach a usable artifact.

Anthropic Claude Code 团队的 Thariq 回顾了一年前最早一批用 Claude Code 制作视频的工作流。当时要得到理想结果，需要漫长迭代并反复纠正视觉细节；如今工具已经从费力的协作制作快速走向更强的生成能力。这也提醒我们：衡量 coding agent 进步不能只看 benchmark，还要看创作者为了得到可用成品，需要投入多少轮反馈。

Guillermo Rauch — Vercel CEO

Vercel CEO Guillermo Rauch warns against accepting AI-generated explanations without understanding them. His example is a viral “performance improvement” attributed to a compiler change even though the AI-written PR description itself said the gain came from altered algorithms and data structures. His broader concern is that a flood of low-quality, unverified prose may devalue reading itself; AI should strengthen human understanding and creativity, not replace verification.

Vercel CEO Guillermo Rauch 警告，不要在没有真正理解的情况下接受 AI 生成的解释。他举例说，一项走红的“性能提升”被归因于 compiler 改动，但 AI 撰写的 PR 描述本身已经说明，收益来自算法和数据结构的变化。他更深层的担忧是：大量低质量、未经验证的文字可能让阅读本身贬值；AI 应增强人的理解力与创造力，而不是取代验证。

Garry Tan — Y Combinator President and CEO

Y Combinator President and CEO Garry Tan shows a production bug-fixing workflow using Capy AI with GStack’s /autoplan and GPT-6 medium reasoning. The substantive signal is that agentic planning is moving into real production debugging: a tool first builds an explicit repair plan, then applies model reasoning to a live issue rather than a toy repository.

Y Combinator 总裁兼 CEO Garry Tan 展示了一个生产环境修 bug 的工作流：Capy AI 配合 GStack 的 /autoplan，并使用 GPT-6 medium reasoning。真正值得关注的信号是，agentic planning 正进入真实生产调试：工具先形成明确的修复计划，再把模型推理用于线上问题，而不是只在示例仓库中演示。

Peter Steinberger — OpenClaw and OpenAI

Peter Steinberger highlights a demonstration he considers clever enough to make the case for AGI feel more tangible. The source post provides no additional textual detail about the quoted demo, so the defensible takeaway is limited: experienced agent builders are still encountering emergent tool behavior that materially shifts their expectations.

OpenClaw 与 OpenAI 的 Peter Steinberger 转发了一项他认为足够聪明、让 AGI 论述更具象的演示。由于原帖没有提供被引用演示的更多文字细节，可确认的结论有限：即使是经验丰富的 agent builder，仍会遇到足以改变其能力预期的新兴工具行为。

Dan Shipper — Every CEO

Every CEO Dan Shipper turned a novelization of Plato’s Protagoras into a short film with Opus 5.5 and published two scenes. The experiment compresses a multi-stage creative pipeline, interpretation, adaptation, screenwriting, and video generation, into a model-assisted workflow, showing how frontier models are becoming production tools for ambitious narrative prototypes.

Every CEO Dan Shipper 先把柏拉图的《Protagoras》改写成小说，再用 Opus 5.5 将其制作成短片，并发布了两个场景。这个实验把理解、改编、编剧和视频生成等多阶段创作流程压缩进 model-assisted workflow，显示 frontier model 正成为复杂叙事原型的生产工具。

PODCASTS

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

The Takeaway: Inception CEO and Stanford professor Stefano Ermon argues that diffusion language models can win on AI inference because they generate many tokens in parallel, mapping computation to GPUs more efficiently than sequential autoregressive models.

Ermon helped originate score-based generative modeling in 2019 and later extended diffusion from continuous media such as images to discrete text and code. A 2024 research result matched an autoregressive model at roughly GPT-2 scale while generating text about 10 times faster, giving him the confidence to found Inception and scale the approach commercially. Its Mercury models now target the quality tier of speed-optimized frontier models while emphasizing much lower latency.

The architectural claim is sharper than “faster output.” Autoregressive inference must produce the next token before starting the one after it, making the workload sequential and memory-bound. Diffusion revises many tokens together, so inference resembles the parallel workload GPUs already handle well during training. That matters for voice agents, coding, real-time interaction, and RL post-training, where rollout generation is often the bottleneck. As Ermon puts it, “the bitter lesson is that the more parallel solution is the one that is eventually going to win.” The bet remains unproven at frontier scale, but it identifies inference economics, intelligence per watt and per dollar, as a more decisive axis than architecture familiarity.

核心结论： Inception CEO、Stanford 教授 Stefano Ermon 认为，diffusion language model 可能赢得 AI inference，原因是它能并行生成多个 token，比逐 token 运行的 autoregressive model 更高效地利用 GPU。

Ermon 在 2019 年参与开创 score-based generative modeling，之后把 diffusion 从图像等连续模态扩展到离散的文本和代码。2024 年的一项研究在约 GPT-2 规模上达到与 autoregressive model 相当的质量，同时文本生成速度快约 10 倍，这促使他创办 Inception 并将技术商业化。其 Mercury models 目前瞄准 frontier lab 速度优化模型的质量档位，同时重点降低 latency。

他的架构判断不只是“输出更快”。Autoregressive inference 必须先生成前一个 token，才能开始下一个，因此负载天然串行且受 memory bandwidth 限制。Diffusion 会同时修订多个 token，使 inference 更接近 GPU 在训练阶段擅长的并行负载。这对 voice agent、coding、实时交互以及 RL post-training 都很重要，因为 rollout generation 往往是瓶颈。Ermon 的原话是：“the bitter lesson is that the more parallel solution is the one that is eventually going to win.” 这一路线尚未在 frontier scale 得到最终验证，但它指出了一个更关键的竞争维度：每瓦、每美元能够获得多少 intelligence，而不是大家对现有架构有多熟悉。

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