AI Builders Digest — 2026-08-22

2026-08-22

AI Builders Digest — 2026-08-22

X / TWITTER

Swyx (AI Engineer / Latent Space)

Swyx highlighted a Latent Space discussion on the model factory behind Thinky-beating models and NVIDIA's reported $6B deal, plus Matt Pocock's /wayfinder, a meta-skill for navigating ambiguous problems through research and targeted “grill” sessions.

Swyx 关注了 Thinky 级模型背后的“模型工厂”及 NVIDIA 据称 60 亿美元交易，还介绍了 Matt Pocock 的 /wayfinder：通过编排研究与针对性追问，帮助开发者在高不确定性问题中找到路径。

Anthropic Claude Code team: Boris Cherny and Thariq

Boris Cherny said Anthropic's upcoming safeguards for “Mythos-class” models were developed to meet enterprise privacy and compliance needs: customers control their data, while Anthropic retains none. Thariq added that Fable safeguards run on customer infrastructure, were developed with roughly 100 companies, and should roll out more broadly in the fall.

Boris Cherny 表示，Anthropic 正在为“Mythos-class”模型准备额外企业安全机制：客户控制数据，Anthropic 不保留数据。Thariq 补充称，Fable safeguards 可运行在客户基础设施上，已与约 100 家企业共同开发，计划秋季扩大推出。

OpenAI Codex and ChatGPT builder Thibault Sottiaux

Thibault Sottiaux clarified that reported Codex usage-limit anomalies often involved unsupported sub2api subscription-to-API traffic, not an undisclosed quota change. He also announced transparent-image generation with GPT-Image-2 and promoted collaborative, shareable ChatGPT Sites.

Thibault Sottiaux 澄清，部分 Codex 用量限制异常来自不受支持的 sub2api 订阅转 API 流量，并非官方暗中调整配额。他还宣布 GPT-Image-2 支持透明背景图片，并展示了可分享、可协作的 ChatGPT Sites。

AI educator Peter Yang

Peter Yang's practical agent hypothesis is that output quality can improve through a manager-agent loop that challenges a worker agent to inspect and retry its work. He also previewed interviews on modern AI evals, viable vibe-coded businesses, ChatGPT Finance and the Grok Bot team's workflow.

Peter Yang 提出，让 manager agent 持续质疑 worker agent、要求复查与重试，可能显著提升输出质量。他还预告了新模型时代的 AI evals、真正形成业务的 vibe-coded 应用、ChatGPT Finance 及 Grok Bot 团队工作方式等访谈。

Meta Senior Director of AI Madhu Guru

Madhu Guru argued that enterprise AI systems need a laddered eval strategy: hill-climbing evals for frontier improvement, regression evals for protecting current behavior, smoke tests for non-negotiable basics and safety, and launch evals approximating real traffic.

Meta AI 高级总监 Madhu Guru 认为企业 AI 系统需要分层 eval 策略：推动能力边界的 hill-climbing evals、保护现有能力的 regression evals、验证基本功能与安全底线的 smoke tests，以及接近真实流量的 launch evals。

https://x.com/realmadhuguru/status/2090595384905113939

Replit CEO Amjad Masad

Amjad Masad emphasized Replit's long-awaited OpenAI partnership and framed the new Free Mode around speed and abundance: lower latency makes coding interactive again, while the free tier can ship substantial projects.

Replit CEO Amjad Masad 强调了与 OpenAI 终于落地的合作，并将新版 Free Mode 的价值归结为速度与供给：更低延迟让编程重新具有交互感，免费模式也足以构建完整项目。

Vercel CEO Guillermo Rauch

Guillermo Rauch described Vercel's direction as “AWS for agents,” while welcoming Bun's simplicity, speed and openness as a strong fit. He also teased a tiny 0.0.5 release with its most-requested feature.

Vercel CEO Guillermo Rauch 将公司的方向概括为“面向 agents 的 AWS”，并认为 Bun 对简单、快速、开放的追求与 Vercel 高度契合。他还预告了体积极小、将加入用户呼声最高功能的 0.0.5 版本。

Box CEO Aaron Levie

Aaron Levie argued that vertical applied-AI companies gain an advantage once they deeply understand a workflow and process enough repeated tasks: post-training can optimize tool use and reasoning for both cost and quality. Proprietary task distributions and high inference costs create room for purpose-built models.

Box CEO Aaron Levie 认为，当垂直 AI 公司足够理解工作流并拥有大量重复任务后，post-training 可同时优化工具调用、推理成本与质量。独特任务分布与高昂推理成本，为专用模型创造了明确空间。

https://x.com/levie/status/2090664811185205722

FPV Ventures partner Nikunj Kothari

Nikunj Kothari explained why venture investors reward extreme ambition: as funds grow, small exits cannot move returns, so firms underwrite early-stage companies against enormous outcomes. The cost of missing a breakout winner can exceed the cost of backing an expensive failure.

FPV Ventures 合伙人 Nikunj Kothari 解释了投资人为何越来越看重“极端野心”：基金规模越大，小型退出越无法影响回报，因此机构按超大结果评估早期项目；错过超级赢家的机会成本可能高于高价投错项目的成本。

https://x.com/nikunj/status/2090585553947517298

South Park Commons GP Aditya Agarwal

Aditya Agarwal said the best founders are reductionists and identified clarity as a core investment attribute: honesty, cutting through uncertainty and charting a path through incomplete information. He used Snowflake CEO Sridhar Ramaswamy as the case study.

South Park Commons GP Aditya Agarwal 认为，最好的创始人是 reductionist，而 clarity 是核心投资标准：诚实面对事实、穿透不确定性、在信息不完整时仍能找到路径。他以 Snowflake CEO Sridhar Ramaswamy 为案例。

Builder Zara Zhang

Zara Zhang shared a compact principle from Claude: “Motivation follows action more than it precedes it.” The builder implication is to use a small committed action to create momentum instead of waiting for motivation.

Builder Zara Zhang 分享了 Claude 的行动原则：“Motivation follows action more than it precedes it.” 对 builder 的含义是：不要等待动力出现，先完成一个足够小但确定的动作来制造动量。

https://x.com/zarazhangrui/status/2090399357145317837

PODCASTS

No Priors: From Restoring Sight to Reimagining the Brain, with Max Hodak

The Takeaway: Science founder and CEO Max Hodak treats the brain literally as a computer, turning some neurological problems from uncertain drug-discovery bets into iterative engineering problems.

Science's near-term product is Prima, a retinal implant paired with laser-projecting glasses that bypasses damaged photoreceptors and stimulates the retina directly. After acquiring the technology from France's Pixium, Science obtained European marketing approval; Hodak says patients have used the restored visual signal to read and complete puzzles. The current product remains narrow-field and monochrome, but the team sees an engineering path toward grayscale and at least red and green.

His broader company design pairs a commercial product with long-horizon research in biohybrid neural interfaces and organ perfusion. The contrarian claim is that BCI is as broad a category as pharma, spanning vision, hearing, balance and motor control. As Hodak puts it: “The brain very literally, very clearly, plainly is a computer.” The deep-tech lesson is to anchor a moonshot in a real business and choose a pathway where each iteration reveals how to improve the product.

核心结论： Science 创始人兼 CEO Max Hodak 将大脑视为一台真正的计算机，从而把部分神经系统疾病从高度不确定的药物发现问题，转化为可持续迭代的工程问题。

Science 的近期产品 Prima 是一枚视网膜植入芯片，配合激光投影眼镜，绕过受损感光细胞并直接刺激视网膜。Science 从法国 Pixium 收购技术后已获欧洲市场许可；Hodak 表示，患者已借助恢复的视觉信号完成阅读和益智题。当前产品仍只有狭窄视野和黑白图像，但团队认为可通过工程迭代增加灰度，并至少实现红色和绿色。

公司的整体设计把商业产品与 biohybrid neural interfaces、器官灌注等长期研究结合。其反常识判断是，BCI 像 pharma 一样是巨大类别，覆盖视觉、听觉、平衡与运动控制。Hodak 的原话是：“The brain very literally, very clearly, plainly is a computer.” 对 deep-tech 创始人的启示是：让 moonshot 锚定真实业务，并选择能通过每轮实验明确下一步改进方向的技术路径。

https://www.youtube.com/watch?v=7HXqMepjvy8

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