P.S. If you liked this thread, you might enjoy my other takes on AI, LLMs, and workflows. Follow me for more experiments, theories, and insights! 🚀

My takeaway: As LLMs evolve, they’re becoming smarter but also less human. Reintroducing older models as a ‘humanizer’ step could help. What do you think? Is this the right direction—or am I overthinking? Let’s discuss 👇

If you’re curious, you can replicate this with modern workflows:
1️⃣ Use a powerful LLM (like GPT-4 or Llama 3.3) to generate a response.
2️⃣ Have an older model (e.g., Llama 1 fine-tunes) rewrite it.
3️⃣ Compare results.
It’s a fascinating experiment!

This small test shows a pattern: Older models trained primarily on human-written text (like Guanaco 65b) can rewrite AI responses with subtlety and natural speech patterns.
This could be huge for applications like emails, articles, or anything requiring a ‘human touch.’

And then there’s Llama 3.3’s attempt to ‘rewrite’ like a human:
Casual? Sure. But a bit messy.

But here’s what happens with GPT-4 or Llama 3.3:
“Honestly, it depends on your goal. Java is great for web apps & enterprise systems, while C# excels in gaming & Windows integration. Both are solid. Any specific type of project you’re looking to work on?”
Polished, but robotic.

Guanaco’s rewrite of Mistral:
“I’d say Java and C# both have their pros and cons. Java’s great for web stuff, Android apps, and big enterprise systems. But C# is amazing for games (Unity), Windows programs, and Microsoft integrations. What are you trying to build?”
Feels more conversational, right?

Mistral’s response:
“Well, it depends on what you're looking to do. Both Java and C# have their strengths. Java is great for web apps and enterprise systems, while C# shines in game dev and Windows apps. What kind of project are you working on?”
Solid, but it screams AI assistant.