Why OpenAI Is Racing to Make Computer Use Superhuman

0:00 – 11:20

From clever demo to something you can actually trust

The conversation opens with a bold claim: computer use is already faster than the average human, and superhuman speed is close. That is a useful provocation because it shifts the frame from novelty to handoff. If the agent can really do the same work you do, faster, the question becomes: what stops you from trusting it? Ari Weinstein starts by saying Astra did not invent computer use from scratch, because OpenAI’s earlier models were already strong. What Astra changed was the part users feel most. It uses fewer thinking tokens to do the same job, so it is a bit faster. It is better at creative tasks, like the launch demo where it talks while drawing in Microsoft Paint. And most importantly, it feels more reliable “around the edges.” Ari gives the kind of distinction only product people obsess over: the difference between 97% reliable and 99% reliable can be hard to describe, but it completely changes whether you trust a system enough to let it run. James Sun adds the enterprise lens. He calls Astra OpenAI’s “most aligned model,” meaning the model is better at following intended goals and resisting manipulation. In plain terms, it is less likely to get tricked by bad instructions hidden in websites or prompts, and less likely to take unintended actions. That matters much more when the work is sensitive, like procurement, vendor communication, or customer interaction, than when it is just playing around in a consumer app. He also mentions a fun benchmark that hints at deeper competence: internally, Astra beat 5.6 at video games, and people have posted examples of it finishing Fallout in roughly 30 hours. The joke is that maybe this skill “translates to something,” but the real point is that navigating messy visual environments under changing conditions is exactly what computer use requires. The contrast between consumer and enterprise use cases runs through the rest of this chapter. On the consumer side, Ari talks about “dots,” where each dot gets its own computer in the cloud and can do things like buy items, book haircut appointments, or handle annoying digital chores. James gives a more personal example of why proactive agents matter. Their internal email system sometimes fails to put external invites on the calendar, so dots now notices the mismatch and asks whether to accept the invite in Gmail. He also uses it to monitor Slack with much more judgment than a brittle automation: flag messages from his reporting chain, identify critical product feedback, open Sentry, pull logs, analyze them, and save him the entire chain of manual triage. That leads naturally into the core product tension: speed, price, and oversight. Ari says speed matters even for background tasks because if the agent is too slow, users think, “I could have just done that myself.” His most important claim is that computer use is already faster than the average person and nearing expert-level speed. James then introduces the hidden cost of that progress: once an agent becomes too fast to watch step by step, you need new forms of control. He points to enterprise controls and a system called Guardian that inspects every action and checks whether it is safe and aligned with the written intent. A consumer may want one-click autonomy. An enterprise may want a dial for human approval, website restrictions, or auditability. The surprising takeaway is that better models do not eliminate product design. They make it more important, because the more capable the agent becomes, the more carefully humans need to decide when to intervene and when to let go.