How to make your own knock-off ChatGPT using GPT‑3 (text‑davinci‑003) — where you can customize the rules to your needs, and access the resulting chatbot over an API.
- Desired prose style can be described in the prompt or demonstrated via examples (neither shown here)
- Answers are generally shorter and factual errors are more common than in ChatGPT
- Generated on text‑davinci‑003 at temperature = 0.7
- Answers are generally shorter and factual errors are more common than in ChatGPT
- Generated on text‑davinci‑003 at temperature = 0.7
I intentionally included an error related to the knowledge-cutoff, where the model confidently asserts Queen Elizabeth II is still alive. Note ChatGPT responds in the same way:
Also note there are typos in the instructions (“requests”) and human input (“Washinton”) and both are ignored, as the intent remains clear. As with ChatGPT, even for severe errors the model almost never asks clarifying questions.
This is just a demo. ChatGPT has a lot of tuning and mitigations to fight abuse and adversarial input, which you don’t get in the DIY approach. You should not expect this to be anywhere near as “safe” as ChatGPT if you deploy it.
@the_venecian If you’re just making a chatbot for your own personal use, though, it’s not that bad. It’s hard to rack up a huge bill just using Playground, so just keep to similar volume and you’ll be fine.
@AravSrinivas @amasad ChatGPT’s behavior in long conversations tends to regress to the mean of how ChatGPT behaves, e.g. forgetting persistent instructions, whereas in Davinci 003 the examples seem to self-reinforce.
@KashPrime @papay0 I haven't tried chat summarization myself, but in this context it seems reasonable to use a recent rolling window of the chat context plus a summary of any transcript at the beginning that needs to be excluded.
Loading suggestions...