twilit flame
GPT-3 doesn't allow you to train the model, but requires you to send along the entire dataset with every prompt, so that it costs more than a dollar every time someone interacts with the AI.
You can't, for example, build a chatbot that acts as support on your website. You can't have it read a dictionary and do translations or definitions based off of that.
What can you do? What's the point of the API? Why don't they allow you to train the model on your own data?
desert pilot
You can fine tune the model for this exact purpose
ionic cosmos
I am also curious about this.
But anyway, if it's not possible, how did they make the Keeper Tax app? https://openai.com/api/
"Helps freelancers automatically find tax-deductible expenses by using GPT-3 to interpret data from their bank statements into usable transaction information."
ionic cosmos
You can fine tune the model for this exact purpose
Can you clarify please?
ionic cosmos
https://beta.openai.com/docs/guides/fine-tuning
Interesting, that answers the tax deductions question... those dont change very often. But what about real-time data that updates once an hour or so? I dont think fine-tuning would work with that, would it? Or if it did, the costs would be astronomical
desert pilot
@ionic cosmos It depends on what kind of data you're receiving, and how much of it relies on a language model.
Regardless you would have to extract and format the information yourself. Training is $0.03 / ~ 750 words so it ultimately depends on what kind of return you're expecting from it
ionic cosmos
<@1066065113238487070> It depends on what kind of data you're receiving, and how...
If it's detailed weather data from every city in the world, updated every 1 minute? Not words but mostly numbers.
desert pilot
That's a pretty massive project. You could save a lot of space and control the noise by graphing it instead of grabbing every piece of data every minute
Do you mind if I ask what you hope to accomplish, assuming you had all this information updating every minute?
ionic cosmos
Do you mind if I ask what you hope to accomplish, assuming you had all this info...
Well, if I wanted to make like a travel advisory app, and have a lot of real-time or near real-time weather condition data for pilots, sailors, other types of travellers?
desert pilot
Unless you wanted to use GPT to make the articles, headlines, advisories etc. Your best bet is to find some sort of weather database API that allows you to do that, and manage all that data yourself
ionic cosmos
Unless you wanted to use GPT to make the articles, headlines, advisories etc. Yo...
Yeah, I was thinking GPT could make human-readable advisories based on real-time data and articles and so on. But getting that real-time data to GPT seems like a costly affair
desert pilot
You'd probably be better off using a graphing software and applying some common technical formulas to catch oddities
You wouldn't feed the real time data to GPT
For example you could set bollinger bands to the graph, and any temperature/wind/whatever that is outside of its bounds you could then feed it to GPT to report on it . There may be a better term, I'm just a stock nerd
But the fundamentals as the stock market I believe are the same... Enough for what you're doing I guess
ionic cosmos
For example you could set bollinger bands to the graph, and any temperature/wind...
Oh, okay. What about stocks? Like if you wanted GPT to give you stock picks or financial advice?
desert pilot
Unless you can find some advantage using language I'm not too sure on that one @ionic cosmos
Maybe from reading the financial reports? There's already sentiment gauges free to use so idk 🤷♂️
ionic cosmos
Maybe from reading the financial reports? There's already sentiment gauges free ...
Hmm ok
ionic cosmos
Maybe from reading the financial reports? There's already sentiment gauges free ...
You dont think there would be a way to use Yahoo stocks API or something in order to give advice based on current trends?
desert pilot
@ionic cosmos There definitely could be. I just haven't looked too far into myself.
Into it myself*
ionic cosmos
Ok
lofty hamlet
You just need to use AI to create your formulas and help you decide how you could present the data to users.
Then store your data in a separate database and use the suggestions from AI to present your data.
The live data shouldn't then need to flow through the AI API
simple crystal
hi
ionic cosmos
You just need to use AI to create your formulas and help you decide how you coul...
But how to connect a separate database to ChatGPT? I havent seen anything about it in the documentation
You say store the data in a separate DB and then use the suggestions to present your data... however, ChatGPT"s responses arent merely based on a formula calculation. If we want the answers to be based on 3 or 5 or 7 year historical daily prices, ChatGPT will need access to that data in order to provide a correct response
random locust