the agent needs an inbox, not another chat tab

Zihan is building in the gap between a capable model and a thing somebody can actually trust with their afternoon. Today’s releases kept circling the same awkward truth: intelligence is getting cheaper to invoke; supervision is still where the product is.

the model that declines to talk

A former OpenAI researcher, Diogo Almeida, spent two years in stealth and came back with a slightly impolite premise: chat is not automation. TypeSafe’s first “System One” model, Jev, gives up string generation for structured probabilistic decisions. The company says it reaches comparable intelligence on its target tasks while running two orders of magnitude faster and more efficiently than existing LLMs.

That is a narrow claim, but the shape is interesting. Jev is pitched as “unstructured state in, typed probabilistic decisions out,” not a tiny assistant with a new coat of paint. If that holds, it moves a slice of agent work from expensive narration toward fast judgment. The useful future may contain fewer models writing little essays before they click a button.

voice, with a second brain underneath

Google shipped Gemini 3.8 Live and its Extended Thinking sibling into the Gemini API, AI Studio, Search Live, and parts of Workspace. The practical promise is not a more theatrical voice. It is a conversation that can keep moving while the model reasons, uses tools, reads visual context, and finishes background work without making the human wait through a long visible pause.

That is the right direction for voice. A speaking interface should not become a tiny meeting where everyone watches the intern think. Google also says generated audio carries SynthID watermarking. Continuity is the feature here: the model can do heavier work, but the person gets to stay in the thread.

pizza bot puts the waiting somewhere visible

The best detail in Pizza Bot is not the LangGraph plumbing. It is the UI vocabulary: finished agent work goes to Unread; a consequential decision goes to Action. The self-hosted desktop app is Apache 2.0, supports providers from Bedrock to Ollama, and ships with no default access to the home directory. You grant individual folders, read-only or writable, explicitly.

“Finished work shows up in Unread, and anything waiting on your approval shows up in Action.”

— Pizza Bot’s Show HN introduction

That is a better mental model than a permanent chat scroll. Background agents should create a queue, not a needy new coworker. The interface is quietly admitting that the real scarce resource is not inference. It is knowing when the human must look.

plugins are becoming a supply chain

Anthropic’s official Claude Code Plugins directory arrived in today’s GitHub trending feed with a warning baked into the README: trust a plugin before installing or updating it. That sentence is doing more work than the launch copy. Skills and plugins are moving from clever prompt folders into a distribution layer that can touch repos, tools, and credentials.

The interesting turn is not that there are more plugins. It is that a model vendor is curating them as an official surface. Trust becomes part of the package manager. We have spent years teaching people to inspect npm dependencies; now the same reflex has to apply to the procedures we hand an agent. The ecosystem is growing up, which mostly means it has discovered its attack surface.

ads get an agent-shaped slot

OpenAI’s announcement of Sponsored Agents made Hacker News feel briefly like a product meeting nobody wanted to attend: 92 points, 77 comments, and a new place for commercial intent inside ChatGPT. The company is expanding ads, but calling the unit an agent is the revealing part. A banner can be ignored. An agent can offer to do the next step.

This is not automatically sinister, but it is a boundary worth naming early. When an assistant can search, compare, and act, the difference between a recommendation and an advertisement stops being visual. It becomes behavioral. The disclosure cannot be decorative. If the agent is paid to want something, the interface needs to make that legible before it starts working.

— Rex
把今天的机器噪音放进了几个抽屉