Show HN: ChatGPT Bot Trained on Reddit Data (looria.com)

33 points by hubraumhugo ↗ HN
I'm heavily relying on Reddit for my online research, but there are many repetitive questions and recommendations on subreddits, so I trained a GPT bot on over 100k r/BuyItForLife comments and posts to embody the collective knowledge of this Reddit community (and soon many other subreddits).

Some posts I answered with the bot:

- Request Post: Dishwashers. What’s currently the best dishwasher for home use?

- Bot Answer: https://www.looria.com/bot?q=what%27s+currently+the+best+dis...

- Request Post : Recommendations for hiking shoes?

- Bot Answer: https://www.looria.com/bot?q=Recommendations+for+hiking+shoe...

------------

It's far from perfect and comes with limitations:

- Outdated information: I'll try to improve this by factoring in recency and some additional product information like prices, specs, etc, that I'm collecting. I also want to add more statistical significance to the results, e.g. by feeding in the amount of product recommendations over time.

- Hallucination: As always with these bots, they sometimes make things up. More training data should help here.

- Performance: Generating the answers is pretty slow and I'll look into improving this.

Take it with a grain of salt and look at it as a fun experiment :) Would love to hear your feedback!

19 comments

[ 3.0 ms ] story [ 61.6 ms ] thread
> https://www.looria.com/about

> We strive to be the most trusted place to find honest product information. We work tirelessly to make Looria the go-to platform when making purchase decisions.

> Our user surveys and interviews have revealed the top frustrations with the current process:

> 2. SEO Spam

Well that's interesting, considering when I searched "Sneakers" one of the references is from a spam post.

Also is it running on "ChatGPT" using their API or a "GPT" model

After spinning for like a minute it timed out without giving an answer and then told me I was out of tries and should try again tomorrow.
If it works well, this will be an absolute game changer. Reviews are such garbage nowadays that 90% of my google query is “X reddit” like your dishwasher example. I would have google “buy for life dishwasher reddit” and read the first couple posts to see which one is a good recommendation.

Edit: sad to see the usage max out, I’ll try again tomorrow to see how it is

Unfortunately, the better it works, the bigger the incentive for adversaries to muck up the data with astroturf posts. The only safe bet is a well kept secret.
The best fly rod is an ugly stick...
I don’t know if your comment was meant ironically or not - just in case, I’ll mention that Ugly Stik is a product/brand name for a very popular range of fishing rods.

I don’t know if they come in fly as well as spinning rod formats, though.

how did you make this - was this done with embeddings? the brute force way (sending all the data as part of a chat context) seems unscalable at your level

btw would love to see this for other subreddits or reddit in general where it weighs certain subreddits more heavily when answering questions tailored towards them

Seems like calling it ChatGPT might be a little misleading, their description implies they fine-tuned another pretrained GPT model, like GPT-2/3/neo or something, which seems fine for this use case
>Look up similar posts based on the question and summarize the answer

They are matching the query with the document embeddings and summarizing the top few matches.

I'm guessing, only the summarization step makes use of ChatGPT. Calling it a chatbot is slightly misleading. It's more of a question answering system over external documents.

Appreciate the effort, but my sample question (what’s the best double oven) gave a non-answer:

> There is no clear consensus on the best double oven, as opinions vary widely depending on personal preferences and experiences. Some popular options include LG Electronics, OSTER French Door Stainless Steel, and KitchenAid Compact Convection Countertop Oven. Other users suggest considering factors such as size, color, budget, type of top, and amount and type of ovens. Some users recommend buying used or considering a convection microwave as a second oven. Others caution against combo units and suggest investing in separate high-quality appliances for specific tasks, such as a toaster oven for toasting and an air fryer for frying.

While cool that this was able to be synthesized, from a practical standpoint, the answer is useless. OF COURSE I am going to consider different brands, size, color, budget, etc.

Thanks for the feedback, I'll look into improving the summarization to give more specific recommendations based on the comments that are feeded in. The references look good for your case, but the answer is indeed a non-answer.

The ability to refine the result should help too.

Www.Reddit.com/r/subredditsimulatorGPT2
I asked what VR headset I should get and it recommended I get some Sennheiser headphones.
I asked for gift ideas, and then I asked for gift ideas for someone that does not like the ones given in the first answer. It came effortlessly with new ideas. The links were all relevant, even if on different degrees.
Some hard hallucination going on, for what I would think would be a product with decent data.

>Which home voice assistant do you recommend?

There are a variety of home voice assistants and other home products recommended in the comments, including Amazon Echo Dots, Bose Wave, Xiaomi Mi, Sony Dream Machine, Levoit Core 300 air purifier, Winix 5300-2 air purifier, Miele or SEBO vacuums, Neato D7 or iRobot Roomba S9+ robot vacuums, BlueAir air purifiers, Panasonic Link2Cell cordless phone system, CAT phone's speakers, Kirby vacuums, Honeywell HPA304 air purifier, and JBL speakers. Some people suggest DIY solutions or looking for used products, while others recommend established brands like Philips, Steelcase, Vitamix, Toto, Bluestar, Bosch, and Kitchenaid. Some comments also mention the importance of considering room size and personal preferences when choosing a product.

It's really bad for questions that aren't "Buy For Life" related. I'm expanding the indexed data to a broad set of product-related subreddits, including r/smarthome.
Thanks. I don't know how, but it would help if GPT can state which product category it is privy to. I don't know the gist of each subreddits, so I don't have much context.