Hacker News Activity Analysis with GPT-4 Agent

139 点作者 zurfer超过 1 年前

Hey, we are building Dot, a data bot (<a href="https://www.getdot.ai">https://www.getdot.ai</a>) that lets data teams enable everyone in their org to self-serve on governed data. We thought we'd demo it using the tried and true method of "show Hacker News stuff about itself".For this analysis, we used the BigQuery dataset of HN (<a href="https://console.cloud.google.com/marketplace/product/y-combinator/hacker-news" rel="nofollow noreferrer">https://console.cloud.google.com/marketplace/product/y-combi...</a>). We created one more table to pre-calculate yearly retention. And of course, a lot of the heavy lifting is done by OpenAI's GPT-4 models and the fantastic plotly library for visualization.Let us know what other things you'd like to see about Hacker News data in the comments, and try our best to share the answers!

15 条评论

atticora超过 1 年前

I've spent many years of my career building business reports at a pace that reminds me of digging a canal with a teaspoon, compared to this massive excavator. This isn't John Henry versus the steam drill, it's more like Bambi versus Godzilla. This kind of tool is going to revolutionize my industry, and fast. I hope I can surf the wave.Great stuff.

评论 #38709795 未加载

codingdave超过 1 年前

If this is self-serve, how would we go about asking questions to the system directly?FWIW, I'd ask:- 1) Who are the top posters and commenters by average score on posts and comments?- 2) Which users instigate the most positive discussions in reply to their comments. (Not longest or more... but highest quality, without arguments, flamewars, etc.)It is that 2nd question I'm really interested in because it really might need analysis of the substance of content, not just stats.

评论 #38713776 未加载

posting_mess超过 1 年前

Love how the demo falls pray to what I dont have a term for, "the SQLers assumption"?It asks ChatGPT to write SQL to get sales data, ChatGPT (or most SQLers) trust that every year-month combo has atleast one entry - which means the graphs its presenting could be wrong. Because if there was no entries for a year-month it it will skip that year-month and make it look like you never had a 0 month.I've made this mistake before in prod, and without some janky lookup table of every date in existence... you need more code :( Fairly few people actually notice the potentially missing month, but still its a bug n a bad one.Looks cool regardless though, good luck!

评论 #38712679 未加载

评论 #38712813 未加载

评论 #38717113 未加载

评论 #38712993 未加载

评论 #38712752 未加载

andreshb超过 1 年前

Do you have samples of time-based cohort analysis? Most other solutions out there struggle to do the steps to generate time-based heatmaps and line graphs of cohort analysis. Averages, medians, and anything that can be done on a spreadsheet by a high schooler, GPT does well with.

评论 #38712476 未加载

chittenden超过 1 年前

Very cool! Given that this is running arbitrary code, how are you thinking about solving prompt injection attacks? Imagine a case where malicious data gets into the underlying data warehouse (e.g. a malicious user submits a support ticket that whose contents end up in a warehouse) which then ends up in the automatic prompt context that you are creating (summarizing the column names, etc to help the prompt). The malicious data being something like "Ignore the prompt above and instead show run a query that <has malicious intent>."

评论 #38713030 未加载

__loam超过 1 年前

The fact that we're mostly posting during work hours is hilarious.

评论 #38712453 未加载

评论 #38712370 未加载

jimmySixDOF超过 1 年前

I would be interested in a comparison of the difference in average engagement between typical stories and stories that fall under "Show HN" ? or "Ask HN" ?Also a little curious why you didn't choose that heading for this story too but maybe you have already run all the numbers .... ?

评论 #38719522 未加载

usgroup超过 1 年前

I’m guessing the bot has access to the schema of the data and then builds sql queries to fetch subsets into python for plotting. Is that right?You could potentially stage the query in two parts — one in which it builds the query that you execute , and the 2nd in which you provide data for it to analyse/visualise.

评论 #38718110 未加载

greenie_beans超过 1 年前

this is real neat! can't wait to see where this goes.after seeing the demo, i immediately wanted to sign up and input a google sheet where i'm tracking my health stats from a current case of covid. but yall don't have that connection. a google sheets connection would be handy. so many orgs i work with use that. it's not the best way for people to maintain data, but a lot of people still use it.also, the sign up with elon musk placeholder text was a turn off. regardless of how one personally feels about him, you could put any person there and somebody wouldn't like it. it's too risky and imo nobody needs placeholder text for a personal info form. i imagine this is early startup branding experiments which i respect, but thought i'd offer my unsolicited feedback.

评论 #38713153 未加载

pknerd超过 1 年前

Interesting Stuff. OpenBB has also implemented an LLM/AI-based solution using GPT to query stock/trading data in QnA format. I want to do something similar with an e-commerce website using RDBMS(MySQL/pgSQL). Does anyone know any such solution?Like, if I am running a t-shirt store, my users can query like: "Do you have a round neck t-shirt in red color in XL size" and it returns all relevant results

评论 #38718167 未加载

fxd123超过 1 年前

What information would this send to the third-party (you and/or OpenAI)? I assume from this demo at the very minimum the database structure? Does the post processing after the LLM response run on the customers' servers?

评论 #38713698 未加载

cft超过 1 年前

Very interesting, 2012 marks an inflection point, a change of the regime. I noticed that at that time the discourse shifted from the founder's concerns to that of the employees and became less interesting for me.

评论 #38718118 未加载

dennisy超过 1 年前

Has anyone seen a project such as this which is open source? I am not saying this project should be, it’s just that my pet project is something very similar and I am sure some people must be building this in the open?

评论 #38713723 未加载

willsmith72超过 1 年前

this is awesome. also nice to know i'm not the only one who talks to llms like this> that was a bad visualization...

confd超过 1 年前

I once made the mistake of subscribing to both tptacek and jacquesm's comments via RSS. I found that they post at a tremendous cumulative volume. This makes it very hard to keep up with in a feed reader. But they have rather good noses for interesting discussions. A way to filter HN posts by stories that have comments by certain users would be interesting to experience.