I have been thinking a bunch about interfaces, how to most optimally communicate information to a user. These new AI tools make this so much easier, but I should question why we even need to show the user anything in the first place? If the models make better + faster decisions than any person could, does showing a person anything make sense at all?
Especially in a time of conflict like agent warfare, will having to wait for a human to make a decision be a cost?
I need to pay close attention to this current obsession with interfaces, am I clinging to human in the loop, will the human get automated out?
For example my recent interest in human x AI interfaces could really be not all that essential, because what happens when these models just get better than people in all aspects, so having a human in the loop at all is a bottleneck!
See! So many little assumptions, dusty corners that you don't notice and are so easy to carry along with you without knowing, like a little stowaway.
YOU NEED TO chill about focusing too much on interfaces.
For the person, or for the agent?
Do you actually want to make an interface for the person? Or for the employee, or rather for the agent!
What does the agent interface look like? How much do they care?
- what is the difference between an interface for an agent and a human?
- what makes something digestible for a human or an agent?
- do humans prefer dense images, symbols?
- agents prefer text, or ideally vectors?
I do not believe that the graph interfaces (agent monitoring systems) being produced are useful or will amount to anything. The current models are not really of the shape where you need to monitor them in real time.
Why I don't think the models need all of this: because you don't really need to steer them, you can force them to ask you a few questions. But overall they will just poke around your codebase and figure out what they need to. I think this is why they make great coding assistants but poor thought assistants (they don't poke around / explore ideaspace with you!!).
See A thinking assistant.
Questions
- What does an interface for swarm monitorability look like? Or even agent monitorability?
- What does a "safety focused" interface look like for AI models, interpretability etc, does seeing it live help?
- What is a warfare based interface like?
- How can you design an interface for fast, optimal decision making (OODA loops?)? Palantir, videogames, etc.
- Given different goals how does the interface differ?
- you want to display all relevant information as quickly and efficiently as possible, minimize ingress to the user
- you want to omit "noise", any irrelevant information that may slow down, or hinder decision making
- you want it to be extremely secure/verifiable if you are in conflict. If your interface gets poisoned or you can damage it in some way, you can damage the decision making + therefore the allocation of resources. Protect the interface itself but also the data sources it pulls from.
Attention
- Should the system monitor the attention of the user?
- Can the same be done for an agent?
- A good interface directs the attention of the user precisely where it needs to be, minimizing noise.
- Detects deviation.
See Will we gesture at all? and Guardian angels.