San Francisco Thinks AI Can Do Everything; Ordinary Businesses Are Still Looking for a Plug
The OpenAI team discusses the real gap in AI adoption: San Francisco thinks models can solve everything, while ordinary businesses haven't even figured out what the tools can do. Computer use may be the closest thing to a 'universal API,' but real-world deployment starts with one invoice, one purchase order at a time.
Living in San Francisco, you hear a lot of ideas like this: the lab can handle everything, so why not just install Codex and let it do the work?
The person saying this is James Sun from OpenAI. He and his colleague Ari Weinstein talked about a specific issue in an a16z interview: the AI adoption gap.
Outside San Francisco, the reality is completely different. For ordinary Americans, ordinary small and medium business owners, and ordinary enterprises, 'understanding what the tool can do' is only the first obstacle. Connecting the tool is the second; deploying it, the third; maintaining it, the fourth.
James Sun and Ari Weinstein both started as founders and later joined OpenAI. Ari works on consumer-facing computer use products, and James works on the enterprise side. Their solution: let AI operate a computer like a person does—clicking, typing, browsing interfaces. This does not require APIs, connectors, or MCP servers.
Ari says the special thing about computer use is its generality. Unlike anything they've built before, where developers need to adapt and someone needs to write connectors, it works with every application.
If there is a complex workflow that requires many different software tools, computer use can string them together and complete the entire end-to-end process. James believes this is still a very large opportunity, and the window is still open.
In their a16z interview, they also discussed these topics: what changes the Astra model has brought to computer use; how consumer and enterprise usage differ; how enterprises can set controls for agents they cannot watch all the time; what agents mean for ad-supported websites; and why the UI will not disappear.
One point is worth noting: the best enterprise use cases are exactly the most boring ones—reconciliation, procurement, invoice processing. This work is not sexy, but it is a real-world pain point.
The full interview is here: https://youtu.be/9v42ZOdiBJY
After a16z posted the tweet, the comment section was more real than the interview itself.
One commenter said that old ERP vendors will soon discover that their ugly UI is their only moat. Another said that San Francisco thinks every company is running agents, while ordinary businesses are still using fax machines and a guy named Dale. The market is huge, really.
But other commenters expressed concerns: on a 2009 ERP interface, speed is meaningless. Every click needs a log, and if you select the wrong supplier from a dropdown menu, it may sit in the books for a month before finance notices. Another person said the adoption gap between San Francisco and ordinary businesses may come from tolerance for the cost of trial and error. Ordinary businesses will only deploy at scale once the net benefit of a single task covers the downside cost.
These responses outline the real world AI deployment has to face: San Francisco is discussing whether agents can speed, while ordinary businesses care about what happens if they choose the wrong supplier. Computer use is indeed the closest solution so far to a 'universal API.' But the real opportunity is not in the lab; it is in those small companies run on fax machines and by a guy named Dale.
发布时间: 2026-10-28 00:01