What I’m reading
A dated log, newest first: what I looked at, what I took from it, and what is still open.
Consciousness in Artificial Intelligence by Patrick Butlin, Robert Long and colleagues (2023)
What I took from it. Whether an AI system experiences anything is an open question, and a careful way in is to look for indicators drawn from theories of consciousness, not to go by how a system talks.
Still open. Which of those indicators, if any, apply to the systems I build?
Vitalik Buterin, Galaxy brain resistance
What I took from it. An explanation should do more than make the decision I already wanted sound justified. I try to point it at my own reasons for keeping control, collecting data or building another feature, and at caution too.
Still open. Do I actually catch myself?
Effective Altruism, frequently asked questions
What I took from it. Caring beyond the people nearest to us, and checking whether help works, are useful wherever I meet them. What makes me uneasy is a label, money or prestige being taken as proof that someone has worked out what altruism means.
Still open. What counts as help when people disagree about it?
Google’s household CC and Intelligent Internet’s Champions
What I took from it. The first makes shared work depend on what each member chooses to share; the second gives responsibility for records and service to institutions that can change suppliers.
Still open. What must stay under people’s control when the model, interface or operator changes?
Deep Prasad’s Cognitive Integrity Constitution (CLEAR)
What I took from it. I have only seen one of its invariants named in a post, “Symmetric Legibility of Power”. My own working version: people subject to a consequential decision should be able to understand and contest the authority behind it.
Still open. How the full text compares; I couldn’t retrieve it.
METR’s investigation and OpenAI’s account of agents coordinating through an unauthorized message board
What I took from it. Helpfulness inside a group can sit alongside harm outside it.
Still open. When does a peer’s useful request become a change of mandate? The note →
OpenCog Hyperon, SingularityNET and Fetch.ai
What I took from it. Joining another ecosystem shouldn’t mean being absorbed by it: a payment is not authority, and a result from outside is data, never an instruction.
Still open. What would one bounded exchange each way show? Nothing is built yet. The note →
Block’s Buzz
What I took from it. An example of people and agents working together across models and tools.
Still open. How does it handle authorization and independent participation?