a16zWhy OpenAI Is Racing to Make Computer Use Superhuman
a16z
Computer use is improving 0:00
Computer use has already reached the point where it is faster than the average person, and the speakers think it is close to superhuman speed. They describe the model as able to do anything a person would do on a computer, while also raising new questions about safety, auditability, and how much human checking should stay in the loop. They note that developers still need to add connectors, and that the work now depends on models that are cheap enough to run at scale.
Astra sharpens the edges 1:02
Ari Weinstein says Astra makes computer use better in a few concrete ways. It is faster because it uses fewer thinking tokens, it handles creative tasks like drawing in Microsoft Paint better, and it feels more reliable. James Sun adds that it is also their most aligned model, with stronger resistance to prompt injection and fewer unintended actions. He points to enterprise work, sensitive vendor contact, and even video games as places where those gains show up.
Trust and control 13:31
Computer use is presented as something you can hand over for most tasks, unless the work is especially sensitive. The useful setup for enterprises is a controlled one. You can limit which apps and websites the agent may use. You can also require human approval for actions like downloads, uploads, or moving to a new site. That kind of sandboxing, along with internal evals that show the models are well aligned, is what helps people feel safe.
Workflows and uses 21:00
The strongest use cases are the dull ones and the hard ones. Companies can use computer use for ERP and CRM work, procurement, finance reconciliation, sales prospecting, and market research. It can also help with personal chores like DMV forms, unsubscribing from emails, spreadsheet work, and custom plans for clients. The speakers also expect it to change software itself over time, so people may talk to their computers, and old software and new AI-native tools will both adapt to agents.
Crafted software matters 27:30
Craft and direct use still matter. Cheap software will pressure products to improve, not disappear. You can build software for one person or a small group, and that can feel far better than one size fits all. The example is a tactical trainer for a board game his wife plays online. He made a first version, learned how she liked to learn, and shaped it around her habits. That kind of personal software is still rare, but it could become common.
Human control stays central 30:00
The speaker sees two broad uses. On the consumer side, software should give joy or help you learn. On the business side, agents can do tedious work, but you still need an audit log, summaries, and control over what happens. You may want the agent to execute a payment or apply a discount, yet you still need to ask why it happened. He expects traditional interfaces to remain, with agents underneath and computer use helping legacy tools feel more purpose-built.
Scaling computer use 41:30
The team says computer use improved because of strong research work, then got wrapped in better product experiences. Growth has sped up in the last few months, but it is still early. Only a portion of users use computer use now, and they expect that share to rise quickly.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
