3 ms·
Ask HN: Is an automated support chatbot powered by GPT-3 a bad idea?
I'm exploring the possibility of creating a support chatbot with GPT-3. With the powerful capacity of GPT-3 to perform well if finetuned with a lot of data, I thought that it could be possible to create a support chatbot powered by it.
That's because support chatbots need to be able to handle a wide range of topics, and GPT-3's ability to generalize from a small amount of data could make it well-suited for the task.
To train the chatbot, I would need a dataset of customer support conversations. Then, fine-tune the chatbot on this dataset.
One challenge with this approach would be that the chatbot might not be able to handle unusual or unexpected customer queries. But this shouldn't be a problem if the dataset is large enough.
I'm not sure if any business would ever use it.
Can you find any reasons why creating a business around this would be a bad idea?
- bell-cot 4y agoHow large, wide, and well-curated a training set do you have for it? How exposed to malicious users will this chatbot be? What's your PR and legal exposure, if the chatbot sometimes says regrettable things?
- saliagato 4y ago1. I don't have any. I'm just exploring the possibility of creating this type of chatbot. 2. It would be as exposed as other GPT-3 applications. What do you mean? 3. Nice question. I would definitely do something about that if I launch the chatbot.
- bell-cot 4y ago2. - My impression is that a lot of "AI"'s and chatbots have coped quite poorly when interacting with malicious users.
- saliagato 4y agoI don't think I get what you are referring to with malicious intent, but OpenAI has some "Safety best practices" [1] that don't seem bad. I'm also sure that they will improve over time. [1] https://beta.openai.com/docs/usage-policies/safety-best-practices https://beta.openai.com/docs/usage-policies/safety-best-prac...
- drdeca 4y agoMaybe you could use it as an autocomplete to make an actual support person able to handle more queries simultaneously? Idk.
- saliagato 4y agoAwesome idea, I wonder if something like this already exists.
- warning26 4y agoThe key problem you're going to run into is when the bot makes promises it can't actually do. For example, a real support tech might say something like "Sure, let me reset your account information." If your bot says that, however, it won't actually be able to do that, and you'll be left with a frustrated customer.