I plan on using this site to test various SaaS ideas.
Feel reach out with suggestions/feature requests; contact: [email protected]
Compared to competitors like "G2A.com":
Finetuned (further training a pre-trained model) OpenAi's GPT3 (ada) on 99196 comment pairs (prompt=parent, response=child) from WSB (July 17 2023 - Aug 3 2023).Filtered comment pairs by the following rules:
The sliding bar sets the temperature parameter (randomness of output - bigger number=more randomness in output returned). Repeatedly running the same prompt will give the same most likely answer at a lower temperature and different answers at a higher temperature (due to increased randomness)
Thoughts the model:
GPT3 wikipedia This example uses Ada which has 350M parameters (weights). Ada is considered old in 2023; it's possible to take an existing popular model with 7B weights from huggingface and finetune that on a rtx3080. That would likely take a few days and be significantly more performant. In comparison, finetuning Ada using Openai's API, took ~40 mins and cost ~$8 usd.
Openai does not want to support finetuning new models (GPT3.5+) anytime in the near future. That might be okay though. GPT4 will support a context window 25,000 words and is expected to have over a Trillion weights. Instead of training a model, you could instead pass in the training data and instruct for a response based on the information and writing style in the data.
How this chatbot performs for sentiment analysis:
If you ask questions like, "Which stock will make me rich", you will almost always get responses like TSLA or NVDA. I think a more useful approach for sentiment analysis would be to feed each comment tree into GPT3.5, and ask it to parse for stock tickers and summarize sentiment - that would give you more straight forward data and it would be easier to find insights
Example prompt: "Which stocks do you own" with 0.65 temperature:
BUD, AMD, NVDA, META, NIO, AWS, GOOGL, SPY, GOOGL, AMZN, BUD, AMZN