Update: My AI chatbot finally made someone laugh on purpose

I've been tinkering with a small language model for about 4 weeks now, mostly trying to get it to tell jokes that actually land. Last night I was testing it with my buddy Mike over Discord. He typed "tell me a joke about a computer" and the bot came back with "Why did the computer break up with the internet? It needed some space." Mike actually laughed out loud and said "that's actually decent." It was just one joke and one person, but after all the times it spat out nonsense or dad jokes that fell flat, that tiny win felt huge. Has anyone else had that moment where your model finally did what you wanted it to do, even just once?

3 comments

3 Comments

drew_jones311mo ago

Dude that space joke is actually solid. I had a similar moment with my LLaMA finetune around week 3 where I asked it to describe a sunset but it started going on about "the sun, tired from its daily climb, finally laid down behind the hills." That was the first time it sounded like a person instead of a textbook. Made me realize the whole trick is catching those tiny wins because 99% of the time the model is just spitting out noise but that 1% feels like you actually taught it something.

kaigibson1mo ago

Wait, is that what success feels like? Because I've been getting that maybe once every few hundred prompts and calling it a win too, haha. Honestly, that's way more poetic than anything I've wrangled out of my model so far. I once got it to describe a rainy day as "water falling from a sad cloud," which I'm pretty sure was just a bug but I printed it out anyway. So yeah, totally with you on chasing that 1% feeling - it's like finding a perfect french fry in a bag of mostly burnt ones. Keep pushing, that sunset line was legit.

morgan_king361mo ago

Have you tried tweaking the temperature or top-k sampling when you're chasing those poetic outputs? I've found that bumping the temp up to like 0.9 and dropping the top-k to 40 gives the model more room to get creative but still keeps it from going totally off the rails. That "water falling from a sad cloud" line is actually great by the way, I'd frame that sucker. The real trick is logging every weird output you like and figuring out what prompt patterns or settings led to it, that way you can start making that 1% feel less random over time. Honestly the best stuff comes when the model surprises you with something you didn't ask for, those are the moments that keep me tweaking.