The point I reached in my earlier post is that, with current techniques, we will be unable to determine whether LLMs are conscious. It is not a helpful question. That’s because questions of consciousness come down to whether an entity experiences qualia. We cannot determine whether entities experience qualia, except through their verbal/non-verbal communication of it. I did not provide an example of an LLM experiencing qualia.
I find myself regularly asking myself, “are LLMs conscious?” I have concluded that this question is unanswerable. To think about it, consciousness must first be defined, and then it must be possible to relate it between different entities. If that is impossible then some method of extrapolating must be found.
Defining Consciousness
Consciousness is not well-defined. There are definitions, of course, but none lend themselves to determining whether something is or is not conscious.
When I started with the Air Force, I was told to push for change where it was needed-to move faster and enable the organization to move faster. The Air Force is a massive bureaucracy designed, in part, to slow change. It was frustrating to push my leadership and senior members to change things, get their pushback, and then be told by those same people to keep pushing.
We rarely make opportunity cost explicit in the military. Because we never think about opportunity cost, risk aversion seems safe.
Opportunity cost - the lost potential gain from an alternative, when a different one is chosen.
Common sources:
Failure to delegate authority
Bureaucratic processes
Delay
Failure to innovate
Recently the Acquisition Transformation Strategy has instructed the military to deliver capabilities to operators earlier - to take on more acquisition risk, in order to buy down operational risk. It is not immediately clear how to judge acquisition risk vs operational risk. Opportunity cost is one way they may be compared. When senior leaders ask us to take acquisition risk to lower operational risk, opportunity cost is the missing yardstick.
I find it fascinating to see how quickly, and in how many directions, practical AI ideas are multiplying right now. I installed OpenClaw recently and set up some daily prompts - show me interesting information, mine some RSS feeds I like - it has helped me learn new things each day… But I did not expect the breadth of different agent orchestration and LLM architectures it would present to me regularly.
I just finished listening to Safi Bahcall’s Loonshots (thank you Libby) because friends kept telling me it was fantastic, and was fundamental to how they think about the innovation we’re all striving towards. They weren’t kidding. The book gave me language for dynamics I’d been managing by intuition, especially during my time with the Shadow Warriors, and showed me how they might maintain success indefinitely.
Bush–Vail Rules and the Artist/Soldier Divide
The Bush–Vail rules hit hard (https://www.infermuse.com/how-to-nurture-loonshots/). They basically say: safeguard your artists (loonshot teams), empower your soldiers (scale teams), and don’t mash them into one bureaucracy. That maps almost perfectly onto how I saw the Shadow Warriors vs. the acquisition command I was embedded in. The Shadow Warriors are artists building weird prototypes, and the acquisition folks are concentrated on keeping the lights on by getting the basics out into the field. I’d sensed mixing those two groups too tightly was dangerous, but Bahcall gave me the structural argument I’d been missing. On the acquisition side I regularly campaigned for less oversight, but the book reminded me there’s a point where “less bureaucracy” can undercut quality (although we can cut a ton of bureaucracy before we reach that point, currently).
A recent Red Hat post about “specification-driven development” caught my interest. I’ve tried bolting AI onto my personal development practices. It doesn’t look like whispering an idea into an LLM then compiling the response… I can’t give the LLM my brain and have it replace all my work beyond the idea.
I have had success when providing a description of the end result, then working alongside an agent to refine and move towards that result iteratively. For me, personally, that’s a significantly different pattern for development than usual.
I remember finding Clapton at the animal shelter near the San Antonio zoo. He was pacing around his small area, it looked like he had a lot of energy. I was looking for a dog who would be, in part, a running buddy. I took a couple dogs out for a walk that first day, but out of all of them Clapton struck me as “the one”.
I came back the following two days and took different dogs for a walk, each time also taking Clapton. At one point there were children playing in one of the dog play areas, and I decided to see how he’d do around them. They wanted to play with him and he was interested in playing with them, it seemed like he liked kids well enough, although he was generally indifferent about people.
Working through a course on “International Security Studies”, I got absorbed in some reading about methods of analyzing international relations. I dove into a rabbit hole and am now able to put my own beliefs and worldview on this into concrete terms.
I’m annoyed by Google’s Analytics. It works great, but it’s heavy and overkill for my needs. Not to mention that it’s very privacy-intrusive. It’s not like I don’t give Google all my data already, but perhaps you don’t make that same choice, and you shouldn’t be forced into it simply by visiting my website.
I’ve been looking for a solution that lets me see what content folks are looking at and where they’re coming from, while being extremely cheap, and easy to maintain. All while reducing the privacy impact. I toyed with building something, but got lost in the “what is the cheapest way I can leverage AWS for this” trade-space.
