June 2, 2026

Peter Yang distilled six takeaways from veteran solo builder Josh Pigford on shipping multiple products with AI agents. The headline lessons: ship early despite the fear ("the idea of spending months working on something before you put it out for other people to use, I think that's a real bad idea"); charge from day one and kill any product that can't cover its own costs; build features in parallel using separate git worktrees (he manages them with Conductor); and have one model review another's code, because "GPT invariably finds three to five bugs that Opus overlooked." Pigford also built a /learnings skill that distills every failed attempt into new rules, and credits 25 years of pre-AI building for how fast he now moves with agents.

Roblox 产品负责人 Peter Yang 总结了资深独立开发者 Josh Pigford 用 AI agent 同时打造多个产品的六条经验。核心要点是：哪怕害怕也要尽早发布（"花上几个月打磨一个东西再拿给别人用，我觉得是个非常糟糕的主意"）；从第一天就收费，养不活自己的产品就果断砍掉；用独立的 git worktree 并行开发不同功能（他用 Conductor 来管理）；以及让一个模型 review 另一个模型写的代码，因为"GPT 总能找出三到五个 Opus 漏掉的 bug"。Pigford 还做了一个 /learnings skill，把每一次失败的尝试提炼成新规则，并把自己如今用 agent 的高效归功于 AI 出现之前 25 年的开发积累。

Guillermo Rauch reported that MiniMax M3 is now the leading open model on the Next.js agent evaluations — landing right behind Opus and GPT-5, but roughly 10x cheaper (and 20x cheaper right now on Vercel's AI Gateway). It's a notable data point for teams weighing open-weight models for agentic coding work on cost grounds.

Vercel CEO Guillermo Rauch 表示，MiniMax M3 已经成为 Next.js agent 评测中排名最高的开源模型，紧随 Opus 和 GPT-5 之后，但成本只有大约十分之一（现在通过 Vercel 的 AI Gateway 更是便宜了二十倍）。对于出于成本考虑、正在评估开源权重模型来做 agentic 编码的团队来说，这是一个值得关注的信号。

Aaron Levie argued that in the era of AI agents, competitive advantage no longer comes from model access — every company has the same models and intelligence — but from how well a business harnesses its institutional knowledge, proprietary data assets, and domain-specific workflows. "The companies that are able to best harness their internal institutional knowledge, existing data assets, and domain-specific workflows -- connected with AI -- will be those that are able to stay ahead," he wrote, adding that the moat will look different in every vertical. Separately, he noted that the AWS-OpenAI partnership likely expands distribution for OpenAI's models while lifting overall token consumption across all providers, given AWS's large committed enterprise contracts.

Box CEO Aaron Levie 认为，在 AI agent 时代，竞争优势不再来自能否用上某个模型——大家用的模型和智能水平都一样——而是来自一家公司能多好地调用自己的机构知识、专有数据资产和特定领域的工作流。他写道："那些能把内部机构知识、现有数据资产和领域专属工作流与 AI 最好地连接起来的公司，才能保持领先。"他还补充说，这种护城河在每个垂直行业里的形态都不一样。另外，他指出 AWS 与 OpenAI 的合作很可能既扩大了 OpenAI 模型的分发渠道，也会因为 AWS 手握大量企业长期合约，而带动所有模型厂商整体的 token 消耗量上升。

Amjad Masad pitched Replit as a full prompt-to-business engine: from a single prompt to a website, a mobile app, monetization, and even a Delaware corporation, with a free tier to start.

Replit CEO Amjad Masad 把 Replit 定位成一个从 prompt 直达"一门生意"的引擎：从一句 prompt 出发，生成网站、移动 app、变现方案，甚至帮你注册一家特拉华州公司，并且可以免费开始。

Thariq has been asking colleagues at Anthropic how they stay in the loop with Claude and fully understand the work being done. He shared a favorite technique from teammate Suzanne — including pairing it with voice mode to make responding easier and more natural — and posted the full prompt as a public gist.

Anthropic Claude Code 团队的 Thariq 一直在问公司同事，他们是怎么在用 Claude 时始终跟上进度、完全理解 Claude 所做工作的。他分享了同事 Suzanne 的一个他很喜欢的做法，包括搭配语音模式（voice mode）来让回应更轻松、更自然，并把完整的 prompt 以公开 gist 的形式发了出来。

Peter Steinberger shared a workflow tweak: he told Codex to ping him whenever it gets blocked and needs his help to proceed — for example, during releases that require npm and are gated behind 1Password. Now he occasionally hears the agent literally talking to him to get unblocked, which he calls "the coolest thing ever."

开发者 Peter Steinberger 分享了一个工作流上的小改造：他让 Codex 在被卡住、需要他出手才能继续时主动呼叫他，比如在需要 npm、并且被 1Password 锁住的发布流程中。现在他时不时能听到 agent 真的在跟他说话、请他帮忙解锁，他形容这"酷毙了"。

Sam Altman highlighted the work of the OpenAI Foundation, saying that helping society become resilient to AI "is going to be incredibly important," with "much more to come here."

OpenAI CEO Sam Altman 着重提到了 OpenAI Foundation 的工作，称帮助社会建立起对 AI 的"韧性"将"极其重要"，并表示"后续还有很多动作"。

The Takeaway: Demis Hassabis has known since the start that the AI race is dangerous, but as the head of just one lab he's concluded that no single company can make the field safe — only governments can, and ultimately a US-China accord.

Sebastian Mallaby, a veteran journalist and author (his earlier books cover venture capital and Alan Greenspan), spent more than thirty hours interviewing DeepMind cofounder Demis Hassabis for his sixth book, "The Infinity Machine." The reflections are worth your attention because almost no one has had this kind of access to one of AI's most important — and most underestimated — figures.

A few things stand out. Hassabis once genuinely hoped to avoid a race between labs ("naive in retrospect"); he now sees safety as a collective-action problem that only government can solve, pointing to the FDA and the UK's AI Safety Institute as proof that public bodies can aggregate real technical talent. Mallaby contrasts Google DeepMind's "let's try everything" culture — "we always make multiple bets, we never really go hard down one avenue" — with Anthropic's concentrated bet on coding, which is why DeepMind was late to both chatbots and coding agents but keeps catching up thanks to Google's bottomless compute and cash. He also reveals "Project Mario," a secret plan to spin DeepMind out of Google backed by a $1B pledge from Reid Hoffman, used as quiet leverage for safety oversight and never executed.

The most surprising thread is spiritual. Mallaby describes Hassabis banging the table about the mystery of why the universe is intelligible: "Maybe if we approach science the right way, we understand more about nature, we will be getting closer to something that we could perhaps call God." And on the rivalry: "Demis has a Nobel Prize. Sam didn't finish his first degree. Therefore, Demis doesn't take Sam very seriously." Mallaby even puts OpenAI's odds of failing by next summer near 50% — not for lack of great tech, but because its business model is brutal against Google's unlimited cash.

要点：Demis Hassabis 从一开始就清楚这场 AI 竞赛是危险的，但作为单独一家实验室的负责人，他得出的结论是：没有任何一家公司能让整个领域变安全，只有政府能做到，而最终需要的是中美之间的某种协定。

资深记者、作家 Sebastian Mallaby（他此前的著作写过风险投资和 Alan Greenspan）为自己的第六本书《The Infinity Machine》花了三十多个小时采访 DeepMind 联合创始人 Demis Hassabis。这些观察值得一看，因为几乎没有人能以这样的深度接触到 AI 领域最重要、却也最被低估的人物之一。

有几点尤其突出。Hassabis 曾经真心希望能避免实验室之间的军备竞赛（"现在回头看是太天真了"）；如今他把安全视为一个只有政府才能解决的集体行动问题，并以 FDA 和英国的 AI Safety Institute 为例，说明公共机构其实能聚集起真正的技术人才。Mallaby 把 Google DeepMind "什么都试一试"的文化——"我们总是同时下多个赌注，从不真正在某一条路上全力压上"——和 Anthropic 在编码上的集中下注做了对比：这正是 DeepMind 在聊天机器人和编码 agent 上都起步偏晚的原因，但靠着 Google 近乎无限的算力和现金，它总能追上来。他还披露了一个代号"Project Mario"的秘密计划：在 Reid Hoffman 承诺投入 10 亿美元支持下，让 DeepMind 从 Google 拆分出去——这被当作争取安全监督权的暗中筹码，但从未真正动用。

最出人意料的是其中关乎信仰的一条线。Mallaby 描述 Hassabis 拍着桌子追问宇宙为何可以被理解这个谜题："也许只要我们用正确的方式去做科学、去更多地理解自然，我们就会越来越接近某种或许可以称之为神的东西。"谈到几位掌门人之间的较量时："Demis 有诺贝尔奖，Sam 连第一个学位都没读完。所以 Demis 并不怎么把 Sam 当回事。"Mallaby 甚至把 OpenAI 在明年夏天之前失败的概率估到接近 50%，原因不是技术不够好，而是在 Google 无限现金面前，它的商业模式实在艰难。