Hacker News new | past | comments | ask | show | jobs | submit login

Are you using GPT-3.5 or GPT-4? A hallucination like FableGL sounds like 3.5, which is not something you should base your expectations of the tech on.

Also, although I haven't seen it before, the Warren Abstract Machine seems like too big of a job for GPT-4. It excels at smaller tasks such as "Convert this C code to idiomatic Rust", or "Write a SQL query that grabs records from X, joins with Y, filters by Z". You might need to make small adjustments manually or by saying "Rewrite it with this in mind".

One really neat trick it did recently is that I uploaded a chart png I found on the internet, and asked it to write Seaborn code that matched the chart's style. It took ~3 follow-up tweak prompts, but then spat out code that worked. It also handles things like "How do I adjust this label that's positioned in chart coordinates over by 10 pixels?", which is an absolute pain to figure out from the docs.

Especially since you're getting answers noting the complexity, I think you're just asking too much of it for now. Try smaller tasks for now, and wait a while to try the big ones again.




So funny SQL query example, do you realize you could describe your exact query to the engine itself in the more precise and short terms? Writing "select X from Y where Z join XX" seems much more practical and rational than using this whole sentence "Write a SQL query that grabs records from X, joins with Y, filters by Z" to achieve the same.


> A hallucination like FableGL sounds like 3.5, which is not something you should base your expectations of the tech on.

So in order to properly evaluate a service, we need to already be subscribed to it in the first place?


Yes. Or just use the api and pay for usage (but in that case It could also get more complicated since you have full control)




Join us for AI Startup School this June 16-17 in San Francisco!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: