|
AI Builders Digest
Saturday, October 31, 2026
|
|
There's a quiet theme running through today's items: who gets to decide what AI is *for*. Microsoft is building fake companies to test its AI. OpenAI is letting real websites opt out of training it. And China's state media is having a cultural crisis because its citizens keep using frontier compute to watch soap operas. These are all, at root, the same argument.
|
|
---
|
|
01
|
China's AI is mostly making soap operas, and the government is not happy about it
|
|
|
A viral claim that 55% of China's AI compute is being spent on short-form entertainment dramas touched a nerve in Beijing. Three state-affiliated outlets published responses in the same week, which is about as close as Chinese media gets to a full-scale freakout. The underlying tension is real: AI-generated micro-dramas are enormously popular (90% of China's 430,000 micro-dramas released this year were AI-produced), but critics call them "braindead" and "shoddily made." The government would prefer its AI investment to signal productive might. The people want melodrama. The people are winning.
|
Why it matters: If you're benchmarking Chinese AI development against American AI development, the token count comparison is almost certainly misleading you. A technology that's eating consumer entertainment time is developing along a very different trajectory than one focused on enterprise productivity or defense. Neither is obviously wrong, but they produce very different capabilities, markets, and blind spots.
|
|
Source →
|
|
---
|
|
02
|
Microsoft built an entire fake company to test Copilot
|
|
|
Microsoft Research published a description of its Industry Evaluation Data Generation Platform, which creates synthetic enterprise environments from public data sources like SEC filings and job postings to test Copilot without touching real customer data. The problem they're solving is real: testing an AI against actual employee emails or internal documents is a privacy violation, but generic benchmarks miss the messy complexity of real business workflows.
|
Why it matters: Every company selling enterprise AI faces this same trap. The environments where the AI needs to work are the ones you're legally forbidden to use for testing. If you're evaluating AI vendors, ask them specifically how they test against your industry's workflows. "We ran it on standard benchmarks" is not the same answer as "we modeled your actual data environment."
|
|
Source →
|
|
---
|
|
03
|
OpenAI will now let websites opt out of AI training through robots.txt
|
|
|
OpenAI announced it will honor a shared robots.txt signal, meaning website operators can add a standard flag to their site and participating AI providers will exclude that content from training data. This is the web's existing "do not crawl" convention, extended to cover training rather than just search indexing.
|
Why it matters: This is meaningful if the other major labs sign on. If Google and Anthropic don't follow, opting out of OpenAI training just means your content trains their competitors instead. Watch for which labs adopt the shared standard and which quietly don't. The ones who opt out of the standard are telling you something about how they think about consent.
|
|
Source →
|
|
---
|
|
04
|
Vercel's CEO wants AI agents to file bug reports on each other
|
|
|
Guillermo Rauch, CEO of Vercel, posted that the future of software engineering is agents communicating with other agents across company boundaries. Vercel is already receiving automated feedback from agents about its platform, and Rauch is encouraging developers to add an AGENTS.md file to their projects so their coding agents know to send Vercel feedback when something breaks.
|
Why it matters: This is a small but concrete preview of what "agents talking to agents" actually looks like in production. Right now it's bug reports. In 18 months it could be contracts, API negotiations, and vendor selection. The infrastructure for machine-to-machine communication at scale is being built piece by piece, mostly by companies whose names you already know.
|
|
Source →
|
|
---
|
|
05
|
The AI Engineer CODE conference is wrapping up a year of coding agent drama
|
|
|
Latent Space's AI Engineer CODE conference is closing out the year with a single-track, invite-only format for what organizers describe as the top 1,000 agentic engineers in the world. The themes: "Scaling without Slop" (managing low-quality AI output), "System One models" (a push toward smarter, cheaper individual models rather than expensive swarms of agents), and ambitious rewrites of major software categories.
|
Why it matters: The "System One models" framing is worth watching. The industry spent 2025 betting on swarms of agents doing parallel work. The argument now gaining ground is that a single model reasoning well beats ten models reasoning poorly and coordinating badly. If that wins out, the cost structure of building AI software changes significantly, and a lot of agent orchestration infrastructure quietly becomes unnecessary.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|