First, the progression of Aaron's hair color continues to break most laws of physics. I'm unsure of how that works.
Also unsure of how an agent doesn't have the same privacy and accountability and security issues that a humans has? When a software engineer borks production, whether through an agent or manually forgetting a semi-colon, these are still the same problems that we've always had? I feel like big enterprise companies are going to do what they always do: Create solutions for problems they've created for themselves and for their customers. It sounds great because their marketing is amazing, but, it's not truly novel; feels like we've heard this story before.
Every time I hear "access controls" I wince a bit because the magic of AI has been really about allowing it to do magic.
And the liability convo is really strange. Software engineering and failures to do it "right" is not the same as a doctor failing on a surgery. The latter has real life consequences and can kill humans while engineering failures can be fixed and rolled back. This type of weird comparative creates a strange mix of FUD and isn't useful for anyone.
The agent identity problem Aaron raises is the one I keep bumping into practically.
'Who is the agent?' sounds philosophical until you realize it determines what the agent will and won't do without being asked. I've been running a personal agent for six months and the biggest unlock wasn't better tools - it was building an explicit operator profile: values, constraints, recurring context, past failure logs.
The agent stopped optimizing for task completion and started optimizing for *my* version of task completion.
The liability point is interesting too - if the creator takes on liability, that's a strong argument for encoding your standards explicitly rather than hoping for alignment through prompting.
First, the progression of Aaron's hair color continues to break most laws of physics. I'm unsure of how that works.
Also unsure of how an agent doesn't have the same privacy and accountability and security issues that a humans has? When a software engineer borks production, whether through an agent or manually forgetting a semi-colon, these are still the same problems that we've always had? I feel like big enterprise companies are going to do what they always do: Create solutions for problems they've created for themselves and for their customers. It sounds great because their marketing is amazing, but, it's not truly novel; feels like we've heard this story before.
Every time I hear "access controls" I wince a bit because the magic of AI has been really about allowing it to do magic.
And the liability convo is really strange. Software engineering and failures to do it "right" is not the same as a doctor failing on a surgery. The latter has real life consequences and can kill humans while engineering failures can be fixed and rolled back. This type of weird comparative creates a strange mix of FUD and isn't useful for anyone.
The agent identity problem Aaron raises is the one I keep bumping into practically.
'Who is the agent?' sounds philosophical until you realize it determines what the agent will and won't do without being asked. I've been running a personal agent for six months and the biggest unlock wasn't better tools - it was building an explicit operator profile: values, constraints, recurring context, past failure logs.
The agent stopped optimizing for task completion and started optimizing for *my* version of task completion.
The liability point is interesting too - if the creator takes on liability, that's a strong argument for encoding your standards explicitly rather than hoping for alignment through prompting.