qiongzhouh

We've looked a bit into CR and Wirecutter, not that deeply into CR yet though. I definitely used Wirecutter for a bit of things in the past, and they have a high level of trust that we'll need to seek to replicate.

I think this is a very good point. We've focused a lot on correctly matching reviews with products / brands, but haven't taken hard enough of a look at astroturfing.

Definitely feel this way too. Sometimes I think to myself that it'd be really great to have an LLM give me a well researched report on say like, recent trends on undisclosed marketing online, wished that we supported that on Lumona, but realized that we'll have to do it eventually, but pretty tough with the current infrastructure

We're working with dumps of Reddit data, which means we don't have to use their API or do any scraping on Reddit itself for now. The data is updated monthly though, so we'll have figure out how to get higher quality data for things that are more time sensitive. We're in contact with the folks at Reddit, so we'll try to see if there are ways to get better data later on.

True, I think I would agree with most of this sentiment. Unfortunately I'm not doing many of these things and am still using my skincare products.

Perhaps we should be surfacing opinions like this beyond just products.

Sorry about that, we didn't make it clear initially that Lumona only has skincare products for now, we'll be working to scale it beyond these products soon, but that message about skincare was probably not clear enough from our post

Wow, thanks for sharing this. I find it interesting that they chose to make it something that I have to wait 1-2 minutes for before I get my AI generated article.

Seems to do a good job for various types of research, will give it a try next time I'm curious about something and need it researched

That's a really good point. I think in our current iteration of the system, if we applied it to subreddits focused specifically on a particular brand, it would not be able to account for the bias there, even if it knows which subreddit the content is from. That's probably too much to ask of the LLM.

We'll have to think about good ways to handle this. Curious about your or others thoughts on these subreddits, how do you process content on these subreddits differently?

Do you have any pointers for learning more about this space? I've personally been pretty skeptical of sponsored products on YouTube, though I find myself getting tempted to / actually trying them out anyway, but haven't thought too hard about small Redditors or Tiktokers getting paid to shill products.

Curious what companies do this / how companies are going about conducting these unpublicized marketing campaigns?

Wow, that voice on the page sounds exactly like what I hear on TikTok / Instagram all the time. Definitely evokes the feeling that I'm about to be entertained by something.

Thanks for sharing that, we'll have to think hard about how to measure the quality / realness of content online beyond the simple things like upvotes and subscriber count

For now, it's just removing AutoModerators and things labeled as bots. Now that I'm reading this again, I realize doesn't really help, since bots pretending to be people recommending products, don't get filtered out.

Sorry for the confusion, we'll try to have general products soon.

I just pushed a change to give an error message explaining what's happening for non-skincare related searches.

I think that's what makes skincare interesting. We want our system to be able to understand your goal, lifestyle, budget, ... and pick out which product is the best for you, given what others who have used the product before said.

With less of this information, the ground truth would probably related to how popular the product is or the average sentiment of people reviewing the product. With this information though, you can compare each one to see which best fits the user's needs. Having compared enough products, you'll eventually figure out which one is the best.

The missing part of the story is when we made an early prototype using GPT-4, leaving it on overnight, and realizing that we've spent several thousand dollars of OpenAI credits...