`This sounds strange but it’s the most clarifying question I’ve found. An agent with no answer to this question has no aesthetic sense, no internal standard it’s protecting. My agent would be embarrassed to surface the same three AI news items everyone else covered. Would be embarrassed to give me angles that don’t fit my brand.
`
I totally get this. It's really similar to how I brainstorm. I have a few agents that are 'toxsec-ified?' and if i pick to write about a linkedin topic, my agent will refuse lol. I really like it though!
Love that your agents refuse you haha that's genuinely the sign they're working 😄
If your agent would happily produce something you'd never post, the persona isn't deep enough yet, IMO. I built similar ones too but I can ALWAYS almost convince it to go for it!! 🤣🤣 Need better guardrails, for sure
Hi @Mia Kiraki - I'm still learning and growing with AI. So, while your new OS looks so intriguing and awesome, I'm trying to figure out if/how I would use the Agents you've created (great, great job by the way!). Any feedback on that would be much appreciated! BTW - I write fiction too:)
This is the article I needed to write, and you've done it so well.
Just like people, agents show us who they are.
But, unlike people, we can fix them to be more in-line with what we intended to build or even realize human judgement and taste isn't in the cards for this particular agent due to the use case.
Wow, love the Harry Potter analogy at play here! I'd agree with you a 100% that most agents are Wormtail. Too much drive to please but no sense of objective and purpose. I feel default agents alternate between Padfoot and Moony haha only if I could 'prong' my way out of them.
Amazing narration, Mia, as always and I'm excited for the Marauders RobotsOS creat next week! 🪄🤖
As a person with a secret life in AI Fine-Tuning (no, I do not bear the Dark Mark), this is absolutely brilliant! (In Ron's voice).
This summarizes exactly how we fine-tune models and what a lot of annotators almost always fail to understand: the agent will do its job come hell or high water. It's actually up to the human to be extra specific about that job, and the Wormtail agent is a perfect comparison: it's fiercely loyal... but to the wrong person. Why? Because its loyalty does not depend on what's right but on what is easy (see what I did there?). And that is where the entire structure collapses.
But then we humans be like: "This agent is broken!" and this is where accountability and responsibility go out of the window, and we become as certain as Minister Fudge, swearing that "He is not back!"
The Wormtail parallel is brilliant, I might need to build an entire AI alignment framework around Marauder archetypes now. You've created a monsteeeeer!!!
P.S. Moony agents just need better guardrails during full moons. We love them anyway dont we? ❤️ 🌕
Love this analogy Mia. And it works so well because as you say without a shared world view being controlled by the human in the loop, AI agents can not achieve what you need, or worse actively sabotage you attempts. My default agent seems to always be Padfoot, which is weird as it is very much not my own personality, although perhaps one I wish I was more like. But at least I haven't fallen through any doorways into the void recently... 😅
I'm absolutely in love with all things Marauders :)
Also, Padfoot agents don't have the worst fate haha! They're super loyal and sharp. Just need a bit of recalibration, maybe. There are definitely worse defaults! 🤣
`This sounds strange but it’s the most clarifying question I’ve found. An agent with no answer to this question has no aesthetic sense, no internal standard it’s protecting. My agent would be embarrassed to surface the same three AI news items everyone else covered. Would be embarrassed to give me angles that don’t fit my brand.
`
I totally get this. It's really similar to how I brainstorm. I have a few agents that are 'toxsec-ified?' and if i pick to write about a linkedin topic, my agent will refuse lol. I really like it though!
Love that your agents refuse you haha that's genuinely the sign they're working 😄
If your agent would happily produce something you'd never post, the persona isn't deep enough yet, IMO. I built similar ones too but I can ALWAYS almost convince it to go for it!! 🤣🤣 Need better guardrails, for sure
yeah it’s honestly a pretty humorous game. watching it balance instructions and its conclusion is entertaining.
It really is! 😅
Wow! I just subscribed and your content runs deep!
Thank you Ana! Subbed back, love yours :)
Hi @Mia Kiraki - I'm still learning and growing with AI. So, while your new OS looks so intriguing and awesome, I'm trying to figure out if/how I would use the Agents you've created (great, great job by the way!). Any feedback on that would be much appreciated! BTW - I write fiction too:)
Hey Ana, let me DM you today! ❤️
This is the article I needed to write, and you've done it so well.
Just like people, agents show us who they are.
But, unlike people, we can fix them to be more in-line with what we intended to build or even realize human judgement and taste isn't in the cards for this particular agent due to the use case.
Some would argue you can fix people too.. 😅 great points Judy!
Wow, love the Harry Potter analogy at play here! I'd agree with you a 100% that most agents are Wormtail. Too much drive to please but no sense of objective and purpose. I feel default agents alternate between Padfoot and Moony haha only if I could 'prong' my way out of them.
Amazing narration, Mia, as always and I'm excited for the Marauders RobotsOS creat next week! 🪄🤖
Hahaha prong your way out of them, that was fun! 🤣 Thank you Raghav ❤️🤗
"Mischief Managed.“
As a person with a secret life in AI Fine-Tuning (no, I do not bear the Dark Mark), this is absolutely brilliant! (In Ron's voice).
This summarizes exactly how we fine-tune models and what a lot of annotators almost always fail to understand: the agent will do its job come hell or high water. It's actually up to the human to be extra specific about that job, and the Wormtail agent is a perfect comparison: it's fiercely loyal... but to the wrong person. Why? Because its loyalty does not depend on what's right but on what is easy (see what I did there?). And that is where the entire structure collapses.
But then we humans be like: "This agent is broken!" and this is where accountability and responsibility go out of the window, and we become as certain as Minister Fudge, swearing that "He is not back!"
Excellent article, Mia, as always.
P.S: At first my agents were like Moony, haha.
Mischief managed indeed haha 🗺️
The Wormtail parallel is brilliant, I might need to build an entire AI alignment framework around Marauder archetypes now. You've created a monsteeeeer!!!
P.S. Moony agents just need better guardrails during full moons. We love them anyway dont we? ❤️ 🌕
Haha! I am glad I can be of service!
A whole framework on Marauders Archetypes would be excellent!
Oh absolutely! They are brilliant! They just need to actually direct that brillance towards actual output!
Right? Not hard at all! 🤣
Love this analogy Mia. And it works so well because as you say without a shared world view being controlled by the human in the loop, AI agents can not achieve what you need, or worse actively sabotage you attempts. My default agent seems to always be Padfoot, which is weird as it is very much not my own personality, although perhaps one I wish I was more like. But at least I haven't fallen through any doorways into the void recently... 😅
Thank you Sam! ❤️
I'm absolutely in love with all things Marauders :)
Also, Padfoot agents don't have the worst fate haha! They're super loyal and sharp. Just need a bit of recalibration, maybe. There are definitely worse defaults! 🤣