Prompting is machine governance. Not conversation. Not suggestion. Every instruction changes the equilibrium of a complex system — and the behavior that emerges is in the interaction between all of them.
Hey Gary, important material about how to manage a proper relationship with a non-human. On a broader level, I think that we must teach the world how it is to work in a hybrid work environment. What you did is make that clear in an empirical situation. Where you went through and were able to bring yourself back in the lead. Not everybody has that level of intelligence and critical attitude. Imagine the millions of people who blindly follow agents’ output. Those outputs will somehow be added to sources that agent will use in the future, building on weak guidance, leading to drift. That is where my thinking goes. It reminds of the work of Iain McGilchrist and his seminal work: The Master and his Emissary (inspired by a story from Nietzsche). I prevent you from diving into that brilliant work by just sharing that Claude is the emissary, but he starts to act as the master. The emissary (Claude) has way more knowledge about facts, etc., but doesn’t understand (its own context). That’s where you acted. We are dealing with emissaries (LLMs, Agents, Robots…) who tend to act as masters because they ‘think’ (look at the ‘thinking’ notification when running analysis) that they know but do not fully understand.
Great points, Paul! Even as I was going to press, I still saw issues with the prompting. I switched from trying to constrain it (don't do x, y or z) to a hierarchical, generative structure:
1) deliver awesome output.
2) Without compromising 1), do the work with an economy of tokens.
If our token burn is any indication, lots of places! Beyond content creation and research, we've used it in hiring, leadership assessments, technical support, finance, legal...
Thank you ☺️ Gary for the reference. The hunan perspective on the Humann Loop is amazing. I love the simplicity in explainingg complexity.. Infusing purpose to agentic environments, that is the deeper challenge. Staying aware even more.
I was mentioning in my response to Paul, experimenting with hierarchical objectives as opposed to exhaustive constraints. I am amazed at the emergent properties that spawn from tiny tweaks. I'm no expert, but part of the problem seems to be the linear construction of thought. If it writes a mistaken assumption early in its thought construction, it becomes "fact". I'll DM you to discuss.
Have had exactly the same experiences.
Hey Gary, important material about how to manage a proper relationship with a non-human. On a broader level, I think that we must teach the world how it is to work in a hybrid work environment. What you did is make that clear in an empirical situation. Where you went through and were able to bring yourself back in the lead. Not everybody has that level of intelligence and critical attitude. Imagine the millions of people who blindly follow agents’ output. Those outputs will somehow be added to sources that agent will use in the future, building on weak guidance, leading to drift. That is where my thinking goes. It reminds of the work of Iain McGilchrist and his seminal work: The Master and his Emissary (inspired by a story from Nietzsche). I prevent you from diving into that brilliant work by just sharing that Claude is the emissary, but he starts to act as the master. The emissary (Claude) has way more knowledge about facts, etc., but doesn’t understand (its own context). That’s where you acted. We are dealing with emissaries (LLMs, Agents, Robots…) who tend to act as masters because they ‘think’ (look at the ‘thinking’ notification when running analysis) that they know but do not fully understand.
Great points, Paul! Even as I was going to press, I still saw issues with the prompting. I switched from trying to constrain it (don't do x, y or z) to a hierarchical, generative structure:
1) deliver awesome output.
2) Without compromising 1), do the work with an economy of tokens.
No problems, but a very simple hierarchy.
Thank you for the “story,” Gary! I am intrigued with how you are leveraging Claude in your business. We should catch up soon!
If our token burn is any indication, lots of places! Beyond content creation and research, we've used it in hiring, leadership assessments, technical support, finance, legal...
I'll DM you.
Thank you ☺️ Gary for the reference. The hunan perspective on the Humann Loop is amazing. I love the simplicity in explainingg complexity.. Infusing purpose to agentic environments, that is the deeper challenge. Staying aware even more.
I was mentioning in my response to Paul, experimenting with hierarchical objectives as opposed to exhaustive constraints. I am amazed at the emergent properties that spawn from tiny tweaks. I'm no expert, but part of the problem seems to be the linear construction of thought. If it writes a mistaken assumption early in its thought construction, it becomes "fact". I'll DM you to discuss.