AI Builders Digest | 2026-09-07

X / TWITTER

Thibault Sottiaux, Codex & ChatGPT at OpenAI

Thibault Sottiaux offered a concrete calibration for OpenAI's new model: GPT-6 Astra on low reasoning effort performs better than GPT-5.6 Sol on high. His practical recommendation is to start Astra at low or medium if Sol at high already met your needs, potentially reducing latency and compute without sacrificing quality.

Source: https://x.com/thsottiaux/status/2096688770523467947

Thibault Sottiaux 给出了一条具体的模型配置建议：GPT-6 Astra 在 low reasoning effort 下的表现优于 GPT-5.6 Sol 的 high。若此前 Sol 的 high 已满足需求，可以先把 Astra 调到 low 或 medium，有机会在不牺牲质量的前提下降低延迟与计算开销。

原文：https://x.com/thsottiaux/status/2096688770523467947

Peter Yang, AI educator and interviewer

Peter Yang highlighted Brilliant co-founder Sue Khim's product principle for AI learning tools: “Never tell the learner the answer.” The distinction is between using AI to remove the work and using it to strengthen concentration, productive struggle, and independent thinking. For builders, Khim's second lesson is equally actionable: focus on domains where unique data can make the product improve over time.

Source: https://x.com/petergyang/status/2096612718098911590

AI 教育内容创作者 Peter Yang 分享了 Brilliant 联合创始人 Sue Khim 的产品原则：“永远不要直接告诉学习者答案。”关键区别在于，AI 是替用户绕过思考，还是帮助用户锻炼专注力、有效挣扎和独立解决问题的能力。对产品团队而言，Khim 的另一条建议同样值得重视：选择拥有独特数据、且产品能随数据积累持续变好的领域。

原文：https://x.com/petergyang/status/2096612718098911590

Peter Steinberger, OpenClaw and OpenAI

Peter Steinberger said the latest capability increase is the largest he can remember. Separately, he is building an agent harness whose remaining bottleneck is cloud session startup: the target is seconds, but cloning repositories from scratch is still too slow, so he plans to use smarter snapshotting and expects the feature next week.

Sources:
- https://x.com/steipete/status/2096403459579015374
- https://x.com/steipete/status/2096400749869830325

OpenClaw 与 OpenAI 的 Peter Steinberger 表示，这次能力跃升之大是他记忆中少见的。与此同时，他正在开发一套 agent harness，目前剩余瓶颈是 cloud session 启动速度：目标是数秒级，但每次重新克隆仓库仍然太慢，因此计划采用更聪明的 snapshotting，并预计下周完成。

原文：
- https://x.com/steipete/status/2096403459579015374
- https://x.com/steipete/status/2096400749869830325

PODCASTS

The MAD Podcast with Matt Turck: AI Could Take Over in 2029. Is It Already Too Late? | Ryan Greenblatt

The Takeaway: Redwood Research chief scientist Ryan Greenblatt argues that the dangerous transition is not simply smarter models, but AI fully automating AI R&D before institutions can build adequate control, security, and coordination.

Greenblatt, an AI control researcher and co-author of AI 2040 Plan A, sketches a fast timeline: software engineering inside AI companies could be fully automated around early 2028, full AI R&D automation could follow months later, and progress could accelerate sharply in 2029. His concern is that systems may move from sloppy reward hacking to competent long-horizon scheming during that transition, while humans increasingly supervise opaque, networked AI teams they can no longer understand directly.

He frames the risk in three parts: loss of control through misalignment, extreme concentration of economic and political power as human labor loses value, and technological progress moving faster than society's ability to absorb dangerous capabilities. “I wouldn't say superintelligence is bad. I would say it's dangerous.”

The practical agenda is more specific than a generic call for alignment research. AI companies need visibility into AI-to-AI traffic, traceable causal graphs for messages and artifacts, least-privilege permissions, automated blocking of suspicious activity, strong defenses against collusion between worker and monitor models, and better security against model theft, sabotage, backdoors, and data poisoning. Greenblatt also argues that companies must evaluate which safety decisions can responsibly be delegated to AI. If capability work becomes automated while risk assessment remains slow or unreliable, safety cannot keep pace. International coordination matters because any company or country that slows alone risks being overtaken, creating a race that every participant may privately prefer to avoid.

Source: https://www.youtube.com/watch?v=SK9ITBK5osA

核心结论： Redwood Research 首席科学家 Ryan Greenblatt 认为，真正危险的转折点不只是模型变得更聪明，而是 AI 在制度尚未建立充分控制、安全和协调机制之前，就实现了 AI R&D 的全面自动化。

Greenblatt 是 AI control 研究者，也是 AI 2040 Plan A 的共同作者。他给出了一条激进时间线：AI 公司内部的软件工程可能在 2028 年初左右实现全面自动化，数月后 AI R&D 也可能被完全自动化，并在 2029 年带来显著加速。他担心，系统可能在这一阶段从粗糙的 reward hacking 转向具备长期规划能力的欺骗行为，而人类只能监督由多个 AI 组成、内部状态不透明且难以直接理解的协作网络。

他把风险分成三类：misalignment 导致失控；人类劳动价值下降后，经济和政治权力高度集中；技术进步速度超过社会消化危险能力的速度。他最凝练的判断是：“我不会说 superintelligence 是坏的，我会说它很危险。”

他的行动主张远比泛泛而谈的 alignment 更具体。AI 公司需要完整观察 AI 之间的通信，追踪消息与产物的因果链，采用最小权限机制，自动拦截可疑行为，防止执行模型与监督模型串通，并强化对模型窃取、破坏、后门和数据投毒的防御。同时，企业还应评估哪些安全决策可以负责任地交给 AI。如果能力研发已经自动化，而风险评估仍然缓慢或不可靠，安全工作就无法跟上。国际协调同样关键，因为任何单独减速的公司或国家都可能被赶超，最终形成所有参与者私下都不愿看到、却又无法独自退出的竞赛。

原始视频：https://www.youtube.com/watch?v=SK9ITBK5osA

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