Loved this YT short from Artificially talking about the hardest part of aligning AI to do what you actually want it to:
Alignment is tricky. Knowing what people want is trickier.
Some people have this locked down, most don’t. Understanding the core problem, and understanding what you and your users really want is leverage.
People that understand this will have a surplus of leverage and advantage over those who don’t.
Yes, and understanding the difference between a want and a need can make a huge difference.
Very true. To make matters worse, it’s very hard for people to even communicate what they want in a way that other people truly understand, let alone have an AI understand it!
I think the trick with AI is to really zoom into the problem you are trying to solve with it and really dial in the context to that problem. I’m curious what the rest of that podcast talked about, but it sounds like she was talking about a very general goal of “AI should always make decisions as if it was a person”.
Facts.
Agreed. It doesn’t seem to be able to handle truly unique and complex problems without a significant amount of guidance.
The models are getting so good now that I literally feel like I’m the bottleneck. The name of the game seems to be clear in your objective, specify constraints and verification tests. I’m still blown away by the fact that this is literally the same kind of details that people need if you are a PM working with a team of engineers.
I think this all goes back to the fact that it’s really hard to explain what you actually want in a way that someone else (or something else) actually understands in the same way you do. Language is a lossy medium. This is a great post and podcast about this topic in the context of doing early product discovery interviews: Customer language 101 - by Rob Snyder
Yes, 100%!
One thing that I’ve noticed is that you don’t even really need the level of detail up-front, it actually seems more efficient to go back and forth with agents- having them dive into code, breakdown constraints, even create suggestions and build the plan themselves simply through communication.
Working with more of your team however changes the dynamic (and increases the value). A designer can provide quick sketches that help you visualize what needs to happen, same with providing valuable context to the agent (with me being the human in the loop).
The more of your team you can include in this process, the faster you get feedback, the higher the value of the work being produced gets. I say this specifically because without a designer, and other parts of your team, I see where AI falls flat.
BTW @Doug_Curtis , the short article you sent over is really awesome. Also loved their followup article How to build the "right" product - by Rob Snyder
Thank you for the share!
Yeah, and when people start using AI, they swap the hard friction of aligning a team for something much easier.
I think this is where successful teams will win.
Went back to this article @doug_curtis, too good. “bitchin’ ain’t switchin’. Lol
Most of these problems are fundamentally a communication and interpretation problem as you explain. Whether you’re working with engineers, AI models, or customers, the bottleneck is often:
We’re building Glare to work on these communication challenges with AI. Managing an agent requires communicating at the right level, similar to managing people.
It’s super interesting how this is all unfolding. I’m ready for a CEO agent, that guides manager agents, that guides PM, designer, and developer agents 
I’d be curious is see if systems end up leaning into this direction
Came across this gem from Dennis Morales Francis- thought it was a way to think about how to capture what people actually want. Diving into customer data reveals a lot.
Step 1: Collect Customer Language
I save every customer email, support ticket, review, and social media comment. Not in a spreadsheet. In a document where I can read actual sentences.Every quarter, I spend two hours reading through them. I highlight sensory words. Visual phrases get one color. Auditory get another. Kinesthetic get a third.
After 100 comments, patterns emerge. One product might attract 60 percent visual thinkers. Another might split evenly. A third might be 70 percent kinesthetic
This feels like AI could be a great way to create loops for this type of activity.
Yes! Data is even more valuable now than ever and AI became that lever!
Loops become even more powerful when you can:
This is also one of the steps to get us to recursive self improvement. Anthropic made a great post about this, even thought it’s model specific, I think it applies to every business: When AI builds itself \ Anthropic