Types of Conversations with Generative AI
nngroup.com
nngroup.com
>Avoid using this type of prompt when working with AI bots like ChatGPT. While search engines perform best with keywords, AI bots need more context to understand what kind of response or output you’re looking for.
Just a few days ago in another thread I interacted with another user who was complaining of how ChatGPT hallucinated an incorrect response for the origin of the term 'bhangmeter'. I got a very accurate response myself[0], and noted that their shared chat was this 'keyword style,' staccato fragment approach[1]. I usually ask ChatGPT questions conversationally, and with that approach I got an accurate result.
I wonder if / suspect that one reason people seem to have such varyingly different impressions of overall response accuracy with ChatGPT comes down to different prompting approaches. If you don't talk to ChatGPT and just bounce keywords off it instead, you're not using it as designed, and it seems to result in less accurate responses. If you talk to ChatGPT like it's a simple and stupid computer search, it responds to you like one. We're kind of forced to personify these tools a bit if we want the best results.
[0]https://chat.openai.com/share/66788bbc-5c2e-4e0d-9a5d-370b2c... [1]https://chat.openai.com/share/8e9caf23-158b-498a-9261-7f257f...
I think there's a fundamental misunderstanding of what these models are all about. I really question/fear how the average person uses it, since OpenAI might be incentivized to push it towards that context.
No evidence, but I feel it's harder to get it to converse about topics, and avoid lists. It seems to love making lists now.
There's a context size, and there's a "comprehension" size. They appear to be completely unrelated, with comprehension being an incredibly small fraction of the context window.
But I have to ask, how do you update your files when iterating on code within the ChatGPT web UI? Do you just copy the code from your browser and manually find and replace the old code chunks in your files?
I ask because my open source coding assistant aider automates away all of that hassle. It lets GPT edit your local files directly. It does a bunch more beyond that, but just the ergonomics of letting GPT edit your files is a huge win. The coding iterations become effortless.
You might want to give it a try if you're using GPT a lot for coding? You do need to provide your own OpenAI API key, so you pay them for token usage as opposed to the all-you-can-eat $20/month ChatGPT web UI.
"At the end of the study, we conducted in-depth interviews with 14 participants."
I wondered something similar while reading, but recognized the choices of explicit gender and detailed inquiry inclusion while intentionally changing the names. These pointed me to the likelihood this was in-house/controlled research. I believe that is the case from the quote above, though I don't see any additional details about how the study was conducted.