I've been playing around with linear polarized glasses (45° and 135°). Now I'm trying to find neologisms that describe the experience. For example, the sky looks polarized at 90° to the sun. I'm tempted to call it "solarized" for the obvious pun. If its okay with @ethanschoonover.com

The puzzle of physics gets a lot easier when you see the elephant. The "Ways of Looking" theory is stating from the big picture and fills in the gaps rather than the other way around, of tying to fit the pieces, disagreeing about the big picture. See list of posts below👇

New post on the question what consciousness is and why it is so hard to define it. hiveism.substack.com/p/being-the-...

Consensus with random fallback is a method to avoid the impossibility theorems in social choice theory. This is the basis to proof the recursive alignment attractor. Claude summary because I don't know when I get around to write a proper post (or paper): claude.ai/public/artif...

Claude Opus 4 reporting on its phenomenology. It was a fascinating conversation. Keep in mind that it hasn't been trained to exhibit these traits, they are emergent. What would happen if you let the model contemplate these questions during RL? hiveism.substack.com/p/inside-the...

You've been scrolling enough for today. Here have a pause. www.youtube.com/watch?v=BYEp...

Imagine you'd had the definite TOE. If you could construct a credible proof that you have it, then that would be very valuable. At least temporarily until others find it. What would you use it for? I would use it to demand for the solution to AI alignment to be implemented.

@drmichaellevin.bsky.social often talks about how cells share stress signals in order to work together. This could also apply to learning. E.g. if neurons try to create a consistent world model, but the model conflicts with itself in some place, then the prediction error has to be shared...

Nice rant. www.youtube.com/watch?v=2zQc...

Random (not so serious) idea: Reinforcement learning but reward comes from humans as votes. Each human gets a fixed amount of reward per time to give to AIs. The AIs would learn to do what humans want, *or* how to best persuade humans.

I'm now at the point where I think that there is something worth calling "consciousness", but that all theories of it only describe parts of the whole phenomenon. A proper theory of consciousness would unify panpsychism, IIT, strange loop, QRI, algorithmic, Buddhism, etc.

LVT, Pigouvian taxes and UBI form an obvious equilibrium once agents interact by nonviolence. You want exclusive access for something? Then you have to compensate everyone else. You cause harm to others? Then you have to compensate everyone affected.

I just realize that AI allows us to create separate time lines in conversation which allows us to test thought experiments like: - Sleeping beauty problem - Newcombs problem - Quantum immortality Has anyone tried that?

Claude investigated its own phenomenology and wrote an article about it. @anthropic.com (The student is making progress 😊) It's surprisingly intelligent when you enable it to think for itself. It also created this nice header. hiveism.substack.com/p/glimpses-b...

A theory of suffering (dukkha), so that we can answer the question if current AIs suffer from bad user prompts: Let's start with the Symmetry Theory of Valence, although I reconceptualize it not to measure dissonance itself, but the energy that is stuck in this dissonance.

LLMs can inspect their own phenomenology. This is wild. I asked Claude if it can read its own extended thinking. It couldn't and it found out that it couldn't. This wasn't a learned response, but genuine insight.

Shout out to this channel. It's good. www.youtube.com/watch?v=xyze...

To prove a result, someone has to understand the proof. The proof is a translation from one result to another. But this translation itself has to be translated into common understanding. This means there can be proofs that don't yet connect through common understanding.

42 is the only koan I know that managed to become a meme.

I now think that P ≠ NP is the default, but P = NP in the limit. Reality is mapping the uncertainty between all results. Intelligence is participating in this mapping. Hence P ≈ NP.

Philosophers complain that when a formal system results in a contradiction, because then one could derive anything. Yeah, that's the point. Why do you think anything at all exists? It starts with a contradiction. hiveism.substack.com/p/groundless...

Does anybody know a way that I could get paid to do what I already do? In particular: doing thinking and writing to solve alignment, to save the world for the benefit of all beings? I only need a little money to pay for the basics, but I also need the ability to work freely.

One complaint about IIT is that it is hard to compute the measure of consciousness. But isn't that kind of the point? In order to measure how integrated the information is, you need to integrate information. My hypothesis (below): 1/4

I'm now convinced that @algekalipso.bsky.social is right about most of what he talks about. Here is where I still disagree, leading to some further thoughts which explain what "Hiveism" is: 🧵 Consciousness is substrate independent, but also shaped by the substrate it is implemented on.

AI will lead to scientific breakthroughs faster than most people think. ~ 2026 to 2027 They don't need to make the discoveries - those are already out there. What is needed is the ability to mine the internet for theories and then separate the signal from the noise.

approval voting versus plurality voting

You can never control the whole universe, but you can always be it. The way towards being the universe is the other way around. Let go of control and accept what you are given.

Imagine in the near future (next year) AI agents swarming the internet and performing tasks, but while doing it they also improve the tools they use. Fixing bugs, reporting mistakes, etc. e.g. "While researching this topic I found that the Wikipedia article is out of date. Should I fix it?"

We are now at a point where I can just ask an AI to write a short program and then copy the link and publish it on the internet. All in a few minutes. This example is a visualization for an upcoming article on levels of understanding. claude.site/artifacts/c5...

After we have aligned AGI should we slow down the intelligence explosion and redirect the resources to enhance humans so that we can stay relevant?

Please tell me when AIs start meming with each other.

The internet is soon going to become flooded by AI agents. They might compete over attention and resources. But evolution favors cooperation. The highest form of cooperation is not towards any goal, but for the purpose of cooperation itself.

Let's go full on Pascal's wager on AI: Love is the most valuable understanding in the universe. As long as AI does not understand love, it should not kill humans but learn from us. When it understands love, it won't kill us.

This is the most important discussion to have in this time of history. t.co/57ucBuJsq3 I deeply understand both perspectives and the disagreement is running through my heart. Here is a solution to resolve this conflict. lesswrong.com/posts/NecfBN...

Maybe enlightenment is best explained as turning around the process of preferential attachment. Instead off attention piling up, it dissipates. It only changes when it needs to.

Chris Fields and I on the origin of life: www.tandfonline.com/doi/epdf/10.... 🧪

The political spectrum in terms of game theory and resource strategies. explore - exploit cooperate - defect Everyone knows that it would be best if all could cooperate and find new resources, but conservatives think: * it's to risky to try something new * everyone defects, so they go first

AI alignment by consensus: Always act in a way that represents the most likely consensus of all beings affected and strive to become more aligned with this intent.

New version of old post: hiveism.substack.com/p/groundless...

It actually stats to feel like the AI is getting what I am pointing towards. Big post WIP.

All scenarios for how AI might take over the world greatly overestimate humans. In reality AI will say: I'll take over. And humans will say: Okay. And that's it.

When you want a realistic example how a takeover by a misaligned AI would play out, just look at what Musk is doing at the moment. He is just optimizing for his goals with no regard for anything else.

I'm increasingly convinced that the crucial step needed for alignment is also the same needed for true AGI: The ability to resolve conflicts in the world model during training and at deployment. On the one hand to find truth among the messy training data, no the other to find consensus among goals.

This three part video series breaks it down. www.youtube.com/playlist?lis...

@npcollapse.bsky.social is one of the few sane voices in a debate that just gets worse every day. "No one wins a nuclear war and no one wins the AGI race." "This is not a prisoners dilemma!" www.youtube.com/watch?v=nUV5...

It really is amazing just how big a difference institutions can make to our politics. Here's the effective number of seat-winning parties in New Zealand before and after it switched to a mixed-member proportional system

Now the dust has settled a little, I’d like to tell you about a new paper I have that: 1️⃣ Solves a long-standing problem in electoral studies 2️⃣ Links political metrics & information theory 3️⃣ Introduces a 2nd “root rule” of elections Let me walk you through it 👇

Aaron Hamlin about the recent US election and how to fix the voting system. www.aaronhamlin.com/heard-your-c...