AI Crispy TalksThe signal in the AI noise

AI news for 2026-11-04

← previous day
03:11
DeepKeep playing the open-source card, DeepSeek is now shipping agent orchestration, not just weights — a signal that the battleground is moving from base models to scaffolding. The router-worker pattern, where a dispatcher fans GitHub and Feishu requests out to parallel worker sessions, is essentially what commercial coding agents monetize, so an MIT-licensed prototype should get forked fast. The halved p95 latency (29.5 → 15.7 minutes) deserves skepticism, though: it comes from DeepSeek's own simulation with no independent benchmark in sight. If you build on agent tooling, this is one to watch rather than cheer yet — the value lies in whether the community validates the orchestration recipe, and Western agent startups may soon face uncomfortable pricing questions if it holds up.
DeepSeek Open-Sources dsh-superagent, a Router-Worker Coding Agent DeepSeek released dsh-superagent, an MIT-licensed research prototype that turns GitHub and Feishu requests into parallel worker sessions. In DeepSeek’s own simulation, p95 task completion fell to 15.7 minutes from 29.5 minutes. The post DeepSeek Open-Sources dsh-superagent, a Route…
01:43
The author leans on a useful framing: professions are bundles of thousands of tasks, and automation consumes them piecemeal, pushing humans up the complexity ladder. In our view the logic holds only partially — history does support task-shifting, but there's no law that residual work keeps getting harder, since today's models are already chewing through the complex judgment tasks once considered safely human. It's also worth noting the source: Foody runs Mercor, which profits from expert labor training AI, so the "humans stay, just on harder work" narrative conveniently flatters his own market. Good conversation starter, but treat the promo framing with a grain of salt.
The closer AI gets to automating an entire profession, the more complex the work gets. Finance alone consists of thousands of individual tasks, each with its own complexity and context. @jefrankle explains why he believes AI will fundamentally change the work humans do on the latest episode of The Frontier. Watch the full conversation ⬇️
00:04
This is a telling data point for anyone tracking how the EU AI Act lands in practice. If a watermark can be switched off via a sampling parameter, it lives in inference defaults rather than in the model weights — compliance as a checkbox, not a property. In our view, answering the gap with a warning instead of a technical block effectively outsources enforcement to developers. Fair to note that text watermarks were never robust anyway — paraphrase defeats most schemes — so the loophole mainly accelerates the inevitable. The real question is whether the Commission audits such exits in December or accepts compliance on paper.
Anthropic's watermark rollout leaves a loophole on two older Claude models Anthropic is expanding its text watermark to four more Claude models to meet the EU AI Act's December deadline. On the older models Claude Opus 4.6 and Sonnet 4.6, developers can bypass the watermark by tweaking a single sampling parameter. Anthropic addresses the loophole …