|
AI Builders Digest
Tuesday, September 29, 2026
|
|
Slow news day on the product side, but the ideas are moving. Jack Clark's newsletter this week covers a Chinese lab running what might be the first public "outer recursive self-improvement loop," and Box CEO Aaron Levie is thinking through what happens when AI agents start optimizing away the friction that entire industries depend on. Those two stories belong in the same conversation.
|
|
---
|
|
01
|
When AI agents choose for you, whole business models break
|
|
|
Box CEO Aaron Levie posted a thread worth reading slowly. His argument: a large chunk of the economy runs on friction. Switching costs, inertia, the fact that canceling a subscription is annoying on purpose. When AI agents start acting in users' best interests and automatically finding better deals, those switching costs collapse. Levie thinks this reshapes competition dramatically until some new equilibrium forms between disrupted incumbents and whatever replaces them.
|
Why it matters: Your cable bill, your insurance renewal, your bank's savings rate, all of these are priced assuming you won't bother to switch. If your AI agent does the switching for you automatically, that pricing model breaks. This is less a technology story and more a business model story. The companies most exposed are the ones whose margins depend on customer laziness.
|
|
Source →
|
|
---
|
|
02
|
Jack Clark's newsletter: a Chinese lab may have started a self-improvement loop
|
|
|
Import AI 474 covers three things worth flagging. First, Zhipu (a major Chinese AI lab) has apparently begun something Clark calls an "outer RSI loop," meaning the lab is using AI to improve AI in a structured, recursive way. Second, Google is testing TPUs in space, which is either genuinely useful for satellite inference or a very expensive flex. Third, there's a philosophical detour through scientist Michael Levin's paper arguing that minds might be patterns from a kind of platonic space, with brains and bodies as hardware that those patterns run on.
|
Why it matters: The Zhipu story is the one to watch. Recursive self-improvement has been an AI safety concern for years, mostly theoretical. A lab actually running a structured outer loop, even a modest one, moves the conversation from philosophy to engineering. Clark is one of the few people who will report on this clearly and without hype.
|
|
Source →
|
|
---
|
|
03
|
OpenAI puts $10M into local news AI training
|
|
|
OpenAI is expanding the Lenfest AI Collaborative and Fellowship Program with $5 million in cash and up to $5 million in software credits. The Lenfest Institute focuses on local journalism sustainability.
|
Why it matters: Local newsrooms are exactly the kind of institution that gets hollowed out when AI makes content generation cheaper. Funding them to adopt AI tools is one answer to that problem. Whether it works, or whether it accelerates the decline by subsidizing further automation, is genuinely unclear. Worth watching what the newsrooms actually build.
|
|
Source →
|
|
---
|
|
04
|
Microsoft Research Asia turns one in Singapore
|
|
|
Microsoft's first Southeast Asia research lab, opened in July 2025, posted a one-year retrospective. The Singapore lab focuses on four areas: next-generation AI models, domain-specific AI applications, AI-native research practices, and regional talent development. The post is heavy on partnership language and light on specific research outputs.
|
Why it matters: Singapore as a beachhead for AI research in Southeast Asia makes geographic sense, given its university infrastructure and government investment in tech. But a one-year anniversary post that leads with vision statements and partnership depth rather than published papers or deployed products is more PR than progress report.
|
|
Source →
|
|
---
|
|
05
|
Google's XPRIZE-winning AI trailer is worth a watch
|
|
|
Google posted the winning short film from the Future Vision XPRIZE competition. The film is called "The Gifted."
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|