OpenAI's Ari Weinstein: computer use is racing past expert humans
Key points
Key takeaways from an a16z Podcast conversation with OpenAI's Ari Weinstein and James Sun, hosted by a16z's Seema Amble and Fabrizio Serafini (October 2026):
Computer use is entering its scaling moment. OpenAI's computer-using agents are now faster than the average human at using a computer and approaching superhuman — faster than expert users — which Weinstein and Sun argue will change how everyone interacts with software.
What Astra unlocks. The Astra model family is faster, more reliable and its most aligned model yet — better resistance to prompt injection and unintended actions — which matters for enterprise deployments handling procurement, vendors and customers.
Enterprise: legacy software without APIs. The strongest enterprise use cases sit on 15-to-20-year-old legacy systems with no MCP connectors: agents navigate ERPs, reconcile financial data and handle procurement by using the interface the way a human does, with no custom integrations.
Harness matters as much as the model. Speed, cost and control trade off; OpenAI's "guardian" auto-approve system uses a smart model to review every agent action, and enterprises can dial human-in-the-loop and auditability to suit their risk appetite.
Dots in practice. Sun describes using dots for proactive work — flagging external invites missing from calendars and pulling Sentry logs to analyse product feedback — because each dot gets its own cloud computer and browser.
The open web and the ad problem. Ad-funded publishers worry agents bypass ad revenue; OpenAI is working with partners on standards such as WebMCP so agents and humans use sites together rather than agents replacing the experience.
Boring tasks are the opportunity. The most valuable early enterprise use cases are boring, structured work — data entry, compliance, calendar and inventory tasks — and the best founders should build on computer use as the universal interface rather than point solutions.
Looking ahead. Agents will increasingly work in the background, collaborate in real time and adapt to users, extending the life of legacy software while accelerating new AI-native applications.
Read more: a16z Podcast on YouTube
