a16zWhy OpenAI Is Racing to Make Computer Use Superhuman
At a glance
WHAT IT’S REALLY ABOUT
Computer-using agents near superhuman speed, forcing new trust controls
- Ari Weinstein and James Sun describe how Astra improves computer-using agents through faster operation, higher reliability, and stronger alignment, which together increase user trust.
- They argue Computer Use is uniquely universal because it can operate any application like a human, avoiding the need for custom integrations such as connectors or MCP servers.
- The discussion emphasizes an approaching “superhuman” speed threshold that could change how people interact with computers by making delegation the default for many tasks.
- For enterprises, the key blockers are safety, control, and auditability, which the speakers say can be addressed through sandboxing, configurable human-in-the-loop, and action-level oversight systems like Guardian.
- They highlight both product-direction challenges (coherency across multiple browsing contexts and states) and ecosystem tensions (open-web monetization and standards like WebMCP to support hybrid human-agent experiences).
IDEAS WORTH REMEMBERING
5 ideasSpeed is the inflection point that could flip delegation into the default.
Ari Weinstein says Computer Use is already faster than the average human and is nearing “superhuman,” meaning faster than an expert user. He argues this would change default behavior from manual execution to delegating most tasks unless you specifically want to do them yourself.
Computer Use’s core advantage is universality: it works with everything.
Ari Weinstein describes Computer Use as “universal” because it operates software the same way a human does, without requiring developers to build connectors or MCP servers. He contrasts this with other tooling where adoption requires integration work.
Astra’s gains are incremental but meaningful: reliability, speed, and alignment drive trust.
Ari Weinstein and James Sun describe Astra as improving Computer Use “around the edges”: faster (fewer thinking tokens), more reliable, better at certain creative tasks (e.g., drawing in Microsoft Paint), and more trusted in practice. James Sun also calls Astra OpenAI’s “most aligned model,” highlighting resistance to prompt injection and unintended actions.
As agents outpace human oversight, configurable autonomy and auditability become essential.
James Sun frames an emerging tradeoff: as agents get too fast to continuously observe, enterprises need controls for autonomy, safety, and auditability. He cites enterprise controls including adjustable human-in-the-loop, and a system called Guardian that evaluates actions for safety and intent alignment.
Proactivity plus cloud computers turns agents into ongoing operational teammates, not just one-off helpers.
James Sun says Dots are useful for proactive, long-running tasks that aren’t identical each time, like detecting calendar invites missed due to email/calendar mismatches and monitoring Slack for important signals. He also describes workflows where a Dot can navigate tools (e.g., logging into Sentry, pulling logs, analyzing them) to reduce end-to-end effort.
WORDS WORTH SAVING
5 quotesWe've gotten computer use now to a place where it's faster than the average human, and I think we're getting really close to a place where it is superhuman, meaning it is faster than an expert computer user, like probably one of us at this table.
— Ari Weinstein
I feel like there’s a little difference that is hard to perceive but really important between something being ninety-nine percent reliable and ninety-seven percent reliable.
— Ari Weinstein
Computer Use uses computers the same way a human does. It makes the agent have all the capabilities you do, um, which I think is so powerful.
— Ari Weinstein
Our business model is not in line with having agents use our product, right? You think about advertisers, publishers whose business model is entirely on advertisement. If the agent's using the website for them, then where's the ad revenue coming from?
— James Sun
I’m really excited about a world where the agent can sort of adapt to you instead of you adapting to it.
— Ari Weinstein
High quality AI-generated summary created from speaker-labeled transcript.
