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...
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.
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.
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.
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.
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.
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.
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...
That means a lot!
This kinda reminds me of this Paper
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.
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.
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.
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.
Correct. It can’t even actually replace.
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.
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.
Can we actually monitor every response? Like all? Across all devices and interactions
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.