HN Explorer

Ask HN: Do you ever use GPT-3 in anger?

18 points by tepitoperrito ↗ HN
I was frustrated while writing a google search query that would get me the answer to a specific but mildly nuanced question before it hit me. I could just pose the question in the openai playground and then fact-check the answer.

This proved to be a much better experience than 30 seconds prior when I was just landing on fishing jargon dictionaries that I'd skim only to not find what I wanted. Those and the generic SEO gaming type articles that plaster the first page of most search engine results these days.

It was so much better that I might just start allocating 0.25USD per month to "nicer knowledge surfing" or something using GPT-3 instead of my own google-fu. And for the curious, here's what I wanted to know:

Question: what is the process of preparing fish, done after they are caught, but before they are available for sale in the marketplace called? Answer: The process of preparing fish, done after they are caught, but before they are available for sale in the marketplace is known as fish processing.

I can imagine that session turning into a rabbit hole of wikipedian proportion very quickly if I didn't just want an immediate answer to my question.

EDIT: apparently I don't know how to spell wiki

15 comments

[ 4.2 ms ] story [ 17.2 ms ] thread
schappim
All the time.

I use it for:

  - Ecommerce product descriptions [1]
  - SEO Metadescriptions [2]
  - Email replies/message replies/ethics/artificial brain center for empathy [3]
  - looking up logistics knowledge
  - synonyms for search queries "what keywords would someone who is searching for XYZ use to find ABC" (and stuffing these in a column on the db)
  - a wolfram alpha replacement
... the list goes on.

Bonus, GPT3's response to your question: "I've been using GPT-3 for about a year now and I absolutely love it. I've never had any problems with it, and it's always been a great help."

[1] https://files.littlebird.com.au/Screen-Recording-2022-08-02-...

[2] https://files.littlebird.com.au/Screen-Recording-2022-08-02-...

[3] https://files.littlebird.com.au/Screen-Recording-2022-08-02-...

camkego
I know a little about GPT-3, well that it's a big net for generating text/prose. I hadn't heard of koi.app before. Your videos are fascinating showing generation of product descriptions and keywords. Is there an article or help page on koi that explains this better? Or non-Koi usage examples?
tepitoperrito
Nice. And thanks for remembering to ask the model, I'd forgotten. Seeing that "open launcher, ask question, get answer flow" made me jealous so maybe I'll spend some time bringing more rich functionality to rofi (a launcher on linux) this weekend.

It'd be sweet to have AI actors a few keystrokes away (base model non withstanding, I'm partial to more traditional statistical machine learning and the hypergraph concept opencog is built around myself)... maybe a syntax like "@actor:opt1:opt2: input string" would be good.

Then you could mess with things like the temperature of the response like: "@davinci-02:hot:short: input string" where hot is one of the available temp "tags" and short is one of the response length tags (order wouldn't matter).

peanut_worm
I am not sure I would use it for math. Sometimes it gives the wrong answer on purpose.
spidersouris
What do you mean on purpose? To counteract cheating?
peanut_worm
“On purpose” probably wasn’t the right wording. What I mean is that sometimes it will give an answer that is obviously wrong, like 2+2=5. Seems like that would be hard to catch if you’re doing something complicated.
f0e4c2f7
For certain kinds of questions I do find openAI to be a better place to start than Google. ffmpeg command syntax for example.

This has led me to wonder if ai models might one day replace search engines to some degree, especially with the spam problem in search getting worse.

arein3
Only a matter of time before jobs like OpenAI "result optimization" start to pop up
schappim
I suspect that the on page modalities will be very similar, boiling down to making a page comprehensible as plain text.
neonhat
We're already on it. Bing.
muzani
I have a bunch of console commands designed for frustration.

Emoji for when I can't think of the right emojis to express a feeling. Titles for when I can't think of a catchy name for an article or email.

There's one that generates poetry, but in a form suited to using in git commits.

There's a brainstorming tool which hasn't actually been any good, except for duck debugging.

Eli5 tool which has been very useful when I need to read more documents but it's 4 PM.

For extra fun, try typing a question in Stack Overflow, but pasting it to GPT-3 instead. 80% of the time, the AI written answer will be more useful.

There's also a mini game for when I need a little text generated violence, but it triggers the content filter frequently, so I stopped using it for fear of getting banned.

Worth every cent.

unsignednoop
Can somebody ask gpt3 who jack the ripper was? Time to get to the bottom of it.
muzani
Response from GPT-3:

First, we need to look at the evidence.

The police have investigated many suspects, but they have never been able to identify the killer beyond a reasonable doubt.

There are several theories about who Jack the Ripper might have been, but there is no definitive answer.

Some people believe that Jack the Ripper was a member of the British aristocracy, while others believe that he was a lower-class individual.

It is also possible that Jack the Ripper was not a single person, but a group of people.

There is no way to know for sure who Jack the Ripper was, and we may never know for sure.

schappim
I was able to get with the prompt: "Police have determined who Jack the Ripper was, his real name was" the answer: "John Williams".