“ “So during the week, the system prompt for Claude 3.5 Sonnet, specifically the artifact section leaked, someone leaked it on X. And one of the biggest discoveries in it was this ant thinking tags that they have in the prompt itself. So it can actually sort of think before it speaks.
I found that super interesting just to see if it like, maybe that’s the improved experience we’re all having with Claude itself. What did you make of this promptly?
Yeah, it’s interesting because we use this exact technique when it comes to like our horse racing AI and stuff like that, where I actually encourage the AI to like, here’s your scratch pad. Here’s where to sum up your things before making a final decision. And I use exactly what they do, like a HTML style tag, which then gets stripped out in the final output.
So this has actually been mentioned in several papers over the time, this idea of providing an AI with a scratch pad. And we actually use a similar technique in SIM theory to get around censorship in models that have it strong. What I always do is I say, in your response, I[…]”
From This Day in AI Podcast: EP69: Fun with kyutai’s Moshi. SimTheory Beta is Here! + Future Assistants, Jul 4, 2024
This material may be protected by copyright.”
EP69: Fun with kyutai’s Moshi. SimTheory Beta is Here! + Future Assistants
This Day in AI Podcast
This Day in AI Podcast
Try SimTheory Beta: simtheory.aiCommunity: thisdayinai.comShow notes: thisdayinai.com/bookmarks/62-ep69—-00:00 – Fun with kyutai’s Moshi28:06 – SimTheory Beta is available: what is new, what we learnt49:04 – RunwayML Gen-3 Alpha52:06 – Is AI in a Bubble?59:52 – Claude Sonnet Prompt Leak for Artifacts1:07:23 – Salesforce’s 1B Parameter Model1:14:14 – Moshi Interrupts Us
Listen on Apple Podcasts: https://podcasts.apple.com/us/podcast/this-day-in-ai-podcast/id1671087656?i=1000661198446
Sent from my iPhone
Recent Comments