Show HN: I made a Slack bot that qualifies your sign-ups using GPT-3 | Hacker News Reader

Show HN: I made a Slack bot that qualifies your sign-ups using GPT-3

qualify.june.so

38 points0xferruccio21 comments

OP here, this was super fun to build. It all started from playing around with Nat Friedman's GPT browser

https://twitter.com/0xferruccio/status/1599014988693180417

Then after having this running for our product for a couple of days or so we decided to give 10 customers access and they loved it! So expanding access now feels great :)

Open articleView on HN

I don't get this. Why do you need GPT-3 enrichment of the provided email addresses when you can get far better information from other, more specialized, services?(e.g. Crunchbase, Clearbit)

What is the value proposition that GPT-3 provides in your solution that is so unique?

The point of this is that as a B2B startup you spend a lot of time Googling information about your sign ups.

The enrichment is not done by GPT-3 per se, we use Google and feed the relevant information from Crunchbase, Linkedin etc.. to GPT3. This allows you to not have to Google things and lets GPT-3 figure out what is the most relevant result within the Google search you'd be running manually.

It's been saving a lot of time for out team - so we just packaged it for others to use :)

For smaller startups Crunchbase and Clearbit api are very expensive - and don't allow for custom searches like "Does this company signing up have a security engineer" or things like that :)

what does it mean to "qualify" a sign-up? I'm really not sure what this app does from the site.

It finds information about people who sign up, saving the effort of the site administrators, I suppose. It allows for... very easy preferential treatment based on arbitrary metrics such as role, worth, etc. you can find an example in the linked Twitter thread.

Or alternatively it allows for resource-strapped companies not to waste a bunch of their time on leads that cannot yield a net-positive relationship for both sides.

Afaict this is just a rephrasing of what I said. Preferential treatment is exactly what happens when you choose to "not waste a bunch of [your] time on leads that cannot yield a net-positive relationship". In effect you're only focusing on the people that you think matter the most.

My assessment probably sounds overly pessimistic, apologies.

Yes it is just a rephrasing. Just didn’t want other readers new to the concept to view this as some form of troublesome discrimination rather than a pretty basic and critical business move to help everyone in a transaction get to a good outcome.

Discrimination is discrimination, "critical business move" or not.

I appreciate your explanation.

Exactly a lot of the time as a startup you google your sign ups to figure out who you should focus your energies on!

Why not put that stuff in a docs or about page?

The page linked tells me nothing.

These types of executions are why I have great fear for the future and AI. You are taking data from a "confident idiot" and parsing as truth. The output of GTP-3 or any of these models is not clean data or truth, but in many cases appears to be truthy. Systems built on this foundation are (at least given current state of the art models) inherently built on bad data (and many times hard to identify as bad, bad data).

explain why 2+2+1-2+1=5 as a math teacher Sure! When we perform arithmetic operations, we follow a set of rules known as the order of operations. These rules dictate the order in which we should perform the operations in an arithmetic expression in order to obtain the correct result.

The order of operations is as follows:

Perform all operations inside parentheses or other grouping symbols first. Perform all multiplications and divisions (starting from left to right). Perform all additions and subtractions (also starting from left to right). Now, let's apply these rules to the expression 2+2+1-2+1. First, there are no parentheses or other grouping symbols, so we move on to the next step.

Next, we perform all multiplications and divisions, but there are none.

Finally, we perform all additions and subtractions, also starting from left to right. This gives us the following:

2 + 2 + 1 - 2 + 1 4 - 2 + 1 2 + 1 3

Therefore, the result of the expression 2+2+1-2+1 is 3.

I hope this helps to clarify the matter. If you have any further questions, please don't hesitate to ask.

This is incredibly cool, especially the prompts[1] where you describe information retrieval as reinforcement learning problem. Am I understanding that right? I had no idea you could do that.

[1] https://github.com/ferrucc-io/gpt-google/blob/57dfd0dc20d614...

Yes! I think people call this k-shot prompting - where the k is the examples you give to the model!

Apps like Jasper.ai do things like:

```

Give me a list of 10 funny intros to a newsletter:

1. funny intro 1

2. funny intro 2

3. funny intro 3

4. {{ai go here}}

```

But is writing this any faster than python and beautiful soup?

Tangent: you may want to consider moving off the .so TLD, I know Notion has had issued with this in the past, with some workplaces blocking .so and at one point losing control of their domain.[0]

[0]: https://www.datacenterdynamics.com/en/news/phishing-complain...

This is creepy AF. Thank God something like Hide my Email exists. I don’t get this need of modern businesses to track every single fart of your users.

Man, the use case is cool but that is just absolutely incredible that ChatGPT can browse the internet. Like holy cow.

This is a thing for b2b sales and lead generation or something of the sort?

I wonder.. could be applied to domain-specific qualifications?

Yes you can customize the prompts and google queries you run like [1] and potentially answer whatever question you can answer by searching Google!

[1] https://github.com/ferrucc-io/gpt-google/blob/57dfd0dc20d614...

another app out of somalia! coming in hot, notion! (the somali top level domain :) )

Reply on news.ycombinator.com