Jev owns System One. Who owns System Two?

Jev makes a good case for automating the quick decisions and paying more attention to who sets them up. A couple of weeks ago, we wrote about the model becoming cheap enough to put almost anywhere. Give it some text and a set of options, and it chooses among them instead of writing a response. TypeSafe calls these System One models, borrowing Daniel Kahneman's distinction between fast, intuitive thought and slower, deliberate System Two reasoning.

That leaves an organizational version of the System Two question: who decides what the options should be, what counts as a good choice, and what context the model needs?

This week, Scott Brinker described the work behind that context. Producing a marketing campaign used to involve people who knew the company, its customers, and what would feel off-brand. Some of that judgment happened in the conversations required to get the work out. An agent that produces the campaign from a short prompt skips those conversations, too.

We should want the faster production. But context can rot as it grows, and giving each agent more documents won't settle a disagreement the people writing those documents haven't settled. Someone still has to choose the goals, resolve the conflicts, and make sure changes reach everything acting on the company's behalf. As individual decisions get cheaper and more numerous, that shared judgment becomes more valuable.

Top Stories#

AI is creating a growing backlog of things to verify. Mathematical results, completed tasks, and generated software all need ways to establish what is correct as production accelerates. OpenAI's mathematics release includes results from an internal frontier model and computer-checkable versions of many proofs in Lean, a programming language for formal verification. OpenAI also plans to fund workshops to help people understand major AI-produced results. Producing a result, checking its proof, and understanding what it means are becoming distinct jobs.

ChatGPT can now build the interface for its answer. OpenAI's GPT-6 rollout introduces Intelligent UI: responses can include buttons, forms, charts, and interactive tools inside the conversation. Ask about retirement savings and you can get a calculator to adjust, rather than another explanation of compound interest. The rollout began October 7 for paid plans, expanding to Free and Go the following day. This makes small, temporary software much more useful: you don't have to decide something deserves an app before getting an interface you can use.

An agent finishing cleanly tells you surprisingly little about whether it succeeded. Microsoft's ThinkingBox evaluation checks the database state agents leave behind. Across 121,680 trials with 12 models, roughly 80,000 failed those checks. About two-thirds of the failed attempts still ended cleanly, used a tool that changed state, and reported no final tool error. These are controlled business-workflow tests, not a failure rate for agents everywhere. But they make a useful distinction: a completed run is an operational fact; a completed job is something you have to check against the job itself.

Top Links#

• EmbeddingGemma 2 (blog.google): Google's new open model puts text, images, audio, and video into a shared representation for search, with local execution that doesn't require sending the material to a cloud service.
• Speech recognition in 16.9 MB (cactuscompute.com): Cactus's Whistle, released last Friday, transcribes seven languages on-device and runs in a browser demo, a pleasant way to discover how little software a useful model can be.
• Training without backpropagation (qlabs.sh): Dust tests an alternative to the standard method of training neural networks, rewarding changes to internal activations that improve predictions; the researchers explicitly say it isn't efficient enough to replace backpropagation today.