X / TWITTER
Sam Altman: safety during model development / 模型开发阶段的安全
OpenAI’s Sam Altman argues that AI safety must address both loss of human control and excessive concentration of power in a person, company, lab, or country. He says OpenAI now prepares explicit safety cases before frontier reinforcement learning runs expected to significantly increase capabilities, extending safety work into development rather than focusing only on model release. He supports consistent federal requirements and independent auditors, while arguing that labs should act before legislation arrives. His position accepts slower progress from the costs of monitoring and safety work, with alignment staying ahead of capability gains.
OpenAI 的 Sam Altman 认为,AI 安全需要同时应对人类失去控制,以及权力过度集中于某个人、公司、实验室或国家的风险。他表示,OpenAI 现在会在预计显著提升能力的前沿 reinforcement learning 训练开始前,明确提出安全论证,将安全工作延伸到开发过程,而不只关注模型发布。他支持统一的联邦安全要求和独立审计,也主张实验室在立法完成前就开始行动。他接受监控和安全工作带来的速度成本,认为 alignment 应当始终走在能力提升之前。
Aaron Levie: safety without weakening competition / 兼顾安全与竞争
Box CEO Aaron Levie supports the safety improvements he attributes to Dario’s proposal, while warning that “pacing” can sound like an arbitrary slowdown or a way to constrain competitors through regulation. He argues that AI’s future role in financial trading, medical devices, biotechnology, defense, and government makes safety and alignment essential. The hard problem, in his view, is meeting those standards without materially slowing innovation or reducing competition.
Box CEO Aaron Levie 支持他所提及的 Dario 方案中的安全改进目标,但也提醒,“pacing” 容易被理解为人为放慢进步,或借监管限制竞争对手。他认为,AI 将用于金融交易、医疗设备、生物技术、国防和政府事务,因此安全与 alignment 必不可少。在他看来,难题是如何在不显著拖慢创新、不削弱竞争的前提下达到这些标准。
Peter Yang: the gap in AI adoption / AI 普及仍有巨大空间
AI tutorial creator Peter Yang argues that AI adoption remains early, citing a chart discussed by Brex’s CEO Pedro: 84% had never used AI, 16% used free chatbots, about 0.3% paid $20 per month, and roughly 0.04% used agents effectively. Yang thinks the chart dates to February 2026 and says adoption has probably increased since then. These are figures from the chart he cites, not a current measurement of global adoption.
AI 教程创作者 Peter Yang 认为,AI 普及仍处于早期。他引用 Brex CEO Pedro 提到的一张图表:84% 的人从未用过 AI,16% 使用免费 chatbot,约 0.3% 每月付费 20 美元,约 0.04% 能有效使用 agent。Yang 认为图表可能来自 2026 年 2 月,并表示此后的采用率很可能已经提高。这些数字来自他引用的图表,并非当前全球使用情况的测量结果。
Peter Steinberger: faster worktrees through folder clones / 用文件夹克隆加速 worktree
OpenClaw and OpenAI builder Peter Steinberger reports that the next release or development channel makes worktrees about 80% faster through folder clones on APFS, Btrfs, XFS, and ReFS. He also reports substantial disk savings; the post does not name the product receiving the change or provide benchmark methodology.
OpenClaw 与 OpenAI 开发者 Peter Steinberger 表示,下个版本或开发通道将通过 APFS、Btrfs、XFS 和 ReFS 的文件夹克隆,让 worktree 操作快约 80%。他还表示,这能明显节省磁盘空间;这条帖子没有明确说明更新属于哪个产品,也未提供基准测试方法。
Zara Zhang: corrections need follow-through / 纠正之后需要继续执行
Builder Zara Zhang reports that Astra sometimes acknowledges a correction without carrying out the replacement action she requested. Her example shows the model agreeing that it should do Y after doing X incorrectly, then stopping instead of proceeding, exposing a gap between accepting feedback and continuing the task.
开发者 Zara Zhang 反馈,Astra 有时会承认纠正意见,却不执行她要求的替代操作。她举的例子是:模型做错 X 后,同意应该改做 Y,却停在那里没有继续,暴露出接受反馈与推进任务之间的脱节。
Boris Cherny: Claude and a historical cipher / Claude 与历史密码
Anthropic Claude Code builder Boris Cherny reports that Fable used Claude to solve the Cyphral Distich, which he describes as a 370-year-old cipher. His post links to the example but provides no details of the solving method or validation.
Anthropic Claude Code 开发者 Boris Cherny 表示,Fable 使用 Claude 破解了 Cyphral Distich,并称这是一道有 370 年历史的密码。他附上了案例链接,但帖子没有介绍破解方法或验证过程。
Nikunj Kothari: startup valuation and employee equity / 创业公司估值与员工股权
FPV Ventures partner Nikunj Kothari warns that higher startup valuations and headline compensation can mask weaker employee-equity outcomes. His post urges candidates to examine market demand, traction, valuation relative to ARR, possible exits, and equity pricing and tax implications, rather than treating a high fundraising valuation as proof of security. He recalls working at a company that raised $60 million and folded three years later, wiping out the equity.
FPV Ventures 合伙人 Nikunj Kothari 提醒,更高的创业公司估值和账面总薪酬,可能掩盖员工股权回报不佳的风险。他在帖子中建议求职者审视市场需求、业务进展、估值与 ARR 的关系、可能的退出结果,以及股权定价和税务影响,不要把高融资估值当成公司稳健的证明。他回忆,自己曾任职的一家公司融资 6000 万美元,却在三年后倒闭,股权价值归零。
通过 Follow Builders 技能生成。