Cerebras says a prototype assistant it built on its own chips booked a dinner reservation in 22 seconds, while three consumer AI assistants it tested took between 4 minutes 36 seconds and 7 minutes 40 seconds, according to a post and chart from Cerebras provided to Lens. The post says a person can make the same booking by hand in 37 seconds. Lens did not view the original post, and every figure below is Cerebras’s own.

The reported times

The chart is titled “The Rise of Personal AI Assistants” and shows the time taken to book a 6 p.m. dinner in San Francisco, split into model and orchestration time, browser and API time, and other overhead.

  • Grok: 7 minutes 40 seconds (3:09 for model and orchestration, 4:31 for browser and API).
  • Cowork: 6 minutes 25 seconds (3:03 and 3:22).
  • Muse: 4 minutes 36 seconds (2:30 and 2:06).
  • Cerebras’s prototype assistant: 22 seconds, with no breakdown shown.

The post says the assistants “took up to 7 minutes and 40 seconds.”

How Cerebras says its prototype worked

  • It built the prototype assistant on Cerebras and tuned it for this task.
  • It checked every restaurant in parallel rather than one after another.
  • It was taught in advance how the booking site works, so it did not have to work that out live.
  • Each call to the model came back in a fraction of the time because it ran on the Cerebras Wafer-Scale Engine, per the post.

The company’s framing

The post argues that consumers are impatient and that assistants will be handed errands once they are faster than people. It points to Google’s experience that delays of a few milliseconds changed how often people searched and how long they stayed, and says, “The real breakthrough will come when agents are faster than we are.” Cerebras says this was one task and that “there is a lot more work to do before agents can handle every errand this well,” and it describes fast inference as a big part of getting there.

What is not stated

The post and chart do not say which booking site or restaurant was used, how many runs were made, or whether the times are averages. They name the three comparison assistants only as Grok, Cowork and Muse, without versions or settings, and do not say how those assistants were configured or whether they had the same advance knowledge of the booking site. They do not identify the model the prototype ran and give no cost figures.

Lens covered Meta’s Muse agent in its post on Spaces and Cerebras’s chip speed claims in an earlier post.

Sources and notes

This post is based on text and a chart from a Cerebras post provided to Lens. Lens did not view the original post, so it could not confirm the author, the date or the underlying test, and the times are unverified. The chart above was redrawn by Lens from the figures shown in Cerebras’s chart; the original graphic is not reproduced. Cerebras makes the wafer-scale chips the prototype ran on.

Image: a restaurant table set for two with a wine glass, orchid and table lamp, by PattayaPatrol via Wikimedia Commons, licensed under CC BY-SA 4.0; cropped. Stand-in photo; it is not related to Cerebras’s dinner-booking prototype assistant and does not show the events described.