Skip to content
Security & Trust Talking point

A Y Combinator speaker says their bots can't tell whose word counts in a Slack thread

A speaker described how Grok Bot, in a Slack thread, lacks context on who to listen to.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
A Y Combinator speaker says their bots can't tell whose word counts in a Slack thread

Photo: Mikhail Nilov / Pexels

In brief
  • A speaker on Y Combinator's show said their bot in a Slack thread lacks context on who the expert is or who decides.
  • Our reading: scoping what a bot can reach does not settle whose request it follows. The new-hire answer covers the first, not the second.

A speaker on Y Combinator's show described a gap in their company's AI bots. In a Slack thread, the bot does not know whom to listen to. The remark came in the episode "The Rise Of AI Coworkers," with Lauren Tan and Roman Ugarte, who built and deployed Grok Bot. The transcript does not say which of them spoke.

What was said

The remark came as the speakers discussed bots joining Slack and meetings. A speaker said that people go back and forth in a thread, and the bot has no sense of who to listen to. It lacks context on who the domain expert is and who has decision-making power.

The speaker compared this to onboarding an employee who has "no EQ and like no like social awareness of the room." Models were not trained for it. The harness, the software wrapped around the model, was not built for it either. The speaker said bots must be taught this at the product, harness and model levels.

A host then suggested the bot could check who its owner is. The speaker agreed that if bots are treated like employees, identity, permissioning, authentication and authorization all need more work. Asked how Grok Bot handles access, a speaker called it an industry-wide question and said, "I don't think anyone has the the silver bullet answer yet."

Their team asks what you would do for a human. Give a new hire credentials scoped to role and tenure. Then, the speaker said, "as they build more trust, you maybe expand their access level."

A speaker also described Quartermaster, a YC project released as open source. Its permissions per Slack channel are those of the user it works with. A general partner gets different access from a software engineer.

Why it matters

Our reading: the speakers treated thread authority and permissions as one identity problem. The new-hire answer, scoped credentials, addresses what a bot can reach. It does not say whose request the bot should follow.

As we understand such bots, a thread reaches the model as text. Unless the system attaches who wrote each message and what that person may order, every line looks alike. A human newcomer reads the room. A bot may not.

A tightly scoped bot could still be steered by the wrong voice inside its scope. Quartermaster's approach is a partial answer, since the bot's reach follows the person it serves. The excerpts do not show how it settles a thread with several people.

For buyers, we would ask vendors two questions. What can the bot touch? How does it decide whose request counts?

The other side

The speaker offered no finished fix. They said no one has the silver bullet. They framed the human comparison as a way not to overthink the problem.

Grok Bot does have some structure. A bot has an owner and a manager, and collaborators can be invited.

The excerpts give no case of a bot obeying the wrong person, and no sign of how often that happens. The remedy described, teaching bots at several levels, is a direction. It was not shown to work.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from