12 Comments
User's avatar
Moscanovski's avatar

The prompt of the day is genius. Thanks Suny. I should have implemented this protocol before my DoscAnovskI experiment. He's reading this, I bet you...

Suny Choudhary's avatar

That means a lot!

Djouadi Abdelmouaiz's avatar

This kinda reminds me of this Paper

Suny Choudhary's avatar

That’s a great reference, thanks for sharing. It ties in really well with this idea that what models say isn’t always a reliable reflection of what’s actually happening underneath. Makes the case for monitoring and validation even stronger.

Kevin Wholley's avatar

You can pick 5 LLM tools use the same prompt and get different results. It all depends on the data they ingested , how they took it in and organized it with the networks it sits on.

Suny Choudhary's avatar

Exactly and that’s kind of the point. The variability itself is the signal that these systems aren’t “truth engines,” they’re probability engines shaped by data, training methods, and architecture. Which is why relying on a single output is risky. That’s also where monitoring and validation come in.

So yeah, the differences aren’t a bug. They’re a reminder that AI outputs need oversight.

Kevin Wholley's avatar

I think we are being led to believe it will solve everything. Then the fear of job loss, and that will happen. It starts to take the form of the next evolution of everything. It should be a tool, but not a replacement.

Suny Choudhary's avatar

Correct. It can’t even actually replace.

Mizieya's avatar

hi Suny how are you - well I know next to nothing about A.I on the level you’re writing about but I found your post really insightful and informative.

Suny Choudhary's avatar

Hey! I’m doing well, thank you! Really appreciate you saying that. That means a lot, especially since the goal is to make these topics easy to understand.

Arvind Agarwal's avatar

Can we actually monitor every response? Like all? Across all devices and interactions

Suny Choudhary's avatar

Great question. In practice, you can’t monitor everything perfectly across all devices and interactions, but you can get very close at the system level.

Most organizations focus on monitoring at key points: inputs, outputs, and integrations (APIs, tools, data access). That’s where the real risk sits.

So it’s less about 100% coverage, and more about placing the right controls where it matters most.