Why OpenAI Is Racing to Make Computer Use Superhuman
- Computer use is projected to transition into a "superhuman" state where agents outperform expert users in speed, potentially leading to a "sea change" in interaction where users prefer agents for most tasks unless manual execution is specifically desired.
- The company plans to shift from manual execution to handing tasks over to agents, leveraging faster models and "less expensive" variations to create more choice while maintaining "coherent" and "understandable" experiences across fragmented browsers and cloud computers.
- Enterprise safety will rely on controls like "Guardian" to allow "auto-approval" of actions based on alignment, with admins using deterministic controls to create "sandboxes" that limit agent access to specific apps and websites while retaining human review of audit logs.
- Legacy software written 10 to 40 years ago will persist for the long term, but computer use will pressure companies to improve compatibility with agents, enabling automation of tedious workflows in ERP, CRM, procurement, and finance without requiring the replacement of existing systems.
- New software classes will emerge where "traditional user interfaces" operate "under the hood" with agents, or where software becomes "personal and unique," designed for specific users or small groups rather than a "one size fits all" approach.
- The diffusion of AI is expected to be slower in "vertical industries" with complex, disparate workflows, creating opportunities for startups to build specialized solutions where "speed and pain" remain a competitive moat.
- Computer use will enable the automation of specific high-volume tasks including financial reconciliation, prospect data collection, video annotation, email management, Slack monitoring, and Sentry log analysis, often where APIs are missing or tasks are time-prohibitive.
- Consumer experiences will expand to include agents buying items, booking appointments, unsubscribing from emails, and managing DMV forms, with voice interaction serving as a primary driver for these automated actions.
- Creative and technical fields such as circuit design, architecture, and hardware engineering will see increased usage as models better understand visual design, while agents are expected to beat games like Fallout in specific timeframes.
- Computer use is scaling at a "YC level of growth week over week" within existing products, with expectations to reach "a billion users" as consumer comfort and trust improve, driven by "cloud" accessibility allowing mobile users to access powerful computers without understanding the mechanics.
- Future interactions will feature "real-time collaboration" and increased "proactivity," where agents adapt to users to identify situations and solve problems without constant initiation, blending multimodal inputs and outputs for fluid experiences.
- Application software companies face challenges regarding "ad impression" revenue when agents visit sites, though adoption of standards like "webmcp" aims to allow agents and humans to use websites together without replacing the human interface.
- The "universal API" for models is expected to become computer use, meaning complex tasks will likely touch computer use at some point in their journey, integrating into all existing products where agentic usage already occurs.
- Software development will evolve to be more "artisan" and "craftsman-like," allowing anyone to build tailored software for niche use cases in a "very short amount of time," alongside new software players that exist alongside legacy users.