Hacker News Reader

Comment by helicone | Hacker News Reader

No worries!

I've seen the same trend with the new AI products in this space, having researched it briefly for this problem. They seem bolted-on and built as afterthoughts. I've been watching a lot of user sessions recently and I've noticed that it's basically impossible to get AI to accurately and consistently classify a problem in a user session by itself. Honestly it seems like AI hurts more than it helps here.

Even so, I'm curious if you would personally find value in something that helps with this bottleneck?

View on HN

> Even so, I'm curious if you would personally find value in something that helps with this bottleneck?

So are you probing HN for a problem you can try to solve, or are you actually trying to help some friends?

If it's the latter, I would have expected YC to have taught them the path to answer, talk to your users

why can't I do both?

if nobody else has the same problem I'm just gonna hack together some scripts, call it a day, and charge them like a contractor.

if the problem is more generalizable then it's worth hunkering down and building something more robust, and charging them like a vendor.

at any rate, part of their problem is people are leaving before they get a chance to talk, and not enough people are talking to them. bit of a catch-22 for them. why not see if the well runs deep?

It's not that you can't do both, but that your post and sequence of comments comes off as misleading. Seeking clarification, because it now sounds like you are more interested in finding a problem to work on

Can you share a link to your friends product? HN could look at it and perhaps give some indication of things that stand out

I didn't mean for it to be misleading, I apologize. In the post I mentioned my concern over the economic viability of solving this problem for them, I intended that to be a clear signal of profit motive on my part.

That's true, I am more interested in gauging if this problem is worth solving in the first place than I am interested in finding a specific solution. I am nonetheless interested in helping my friends.

No, sorry, I can't share a link.

I would find it _very_ valuable.

But I think this is also the hardest task of a PM, so I am skeptical. There is a lot less learning and training material for an AI to use (compared to for example writing code) - so it is no surprise that AI in its current state does often not lead to great results.

that's interesting to hear, thanks! and it does confirm my suspicions.

If you don't mind me asking, when you watch sessions, do you have a better way to prioritize the session you watch than just picking at random?

For example, do you have some way to pull out a bunch of session groups from the data automatically so you only have to watch one session in the group to know what the problems were for all of the sessions in that group?

Sessions are just one "channel" to build up an understanding of the problem, so I try to prioritize the ones where I have prior indication that they might be interesting (churn, complaints in a conversation or feedback via some other channel). But sometimes it can also be interesting to just "surf" through some random ones.

I'm interpreting from this that you have more data than you have time to process, so you go after the signals in order of how close they are to the user's complaints (user convos, etc).

And this implies to me that your ideal scenario is one in which the amount of data coming in from user convos, surveys, complaints, tickets, etc is equal or greater than the amount of time you have as a team to process it, such that you can focus on that and keep yourself productively at capacity.

But what if the amount of such high-priority signals is much higher than what you can deal with? Is it worth clustering that to get a smaller list of actionable trends?

Furthermore, if this is the highest quality data, is there even any need to go in and 'process' ALL of the sessions and bin them regardless of their high-priority signal status? Am I reading you right?

Thanks again for your feedback!

Just to wrap this back, I suspect there's a way to process these sessions really quickly to figure out what the user trends are in ~real-time. This way you can do the same thing you're already doing, but with much more context.

Adding another two channels, with these channels being: 1. A weighted map of stats for all of the sessions at once. Weighted by heuristics you choose or by good defaults. 2. Reports detailing all the natural problem groups the sessions fall into. With breadcrumb trails available for deep-dives.

Most importantly I think there's a way to do this without having to rely on llms at all; by modeling the whole thing as a set of graph problems.

If this sounds reasonable to you lmk.

Have a good one!

Reply on news.ycombinator.com