from Hacker News

Ask HN: Are there any LLM summary features that don't suck?

by forcer on 11/23/24, 7:20 PM with 1 comments

Every tech company and their uncle seems to be rolling out AI features these days, especially ones that promise to summarize all the stuff you’ve got stored in their platforms. From Gmail and Slack to smaller players like Gitbook, everyone’s jumping on the bandwagon—usually behind a shiny premium paywall, of course.

Since we rely on a mix of these big and small tools, we decided to put their new AI toys to the test. We asked them to summarize emails, messages, posts—you name it. The verdict? Meh.

The summaries they spit out are all over the place. They either miss the key points entirely or focus on the most random, irrelevant details. It’s like asking someone for a quick recap of a meeting, and they tell you what snacks were on the table.

So now I’m left wondering: is it just me? Or is Google, Slack, and the rest of the AI crew genuinely struggling to make these features work the way they’re hyped up to?

  • by JohnFen on 11/23/24, 8:27 PM

    I think they're about 75% accurate or thereabouts. Far too much error to be able to rely on them blindly, but accurate enough to be able to be useful as a starting point to aid a human writing up notes or a summary.