Simon Ouellette

Google search volume for "OpenClaw" has experienced a massive crash, plummeting right back down to pre-launch baseline levels almost as quickly as it spiked.

Ok, I think my experiment leaving AI working on stuff 24/7 ends here. It doesn't work. Code explodes in complexity, results are not that great, the AI can't get past hard walls (it is still completely unable to even *grasp* SupGen), and it is insanely expensive (spent ~1k over the last 2 days). The best results are on the JS compiler, mostly because it is familiar (compared to inets), but not worth losing control over the codebase. I think the dream of having AI's working on the background and making real progress on things that matter (i.e., truly new things) isn't here yet. It is still a machine hard-stuck on its own training data, incapable of thinking out of the box. It is great for building things that were already built. But not new things Also coding normally has the under-appreciated advantage that you're doing two things at the same time: building a codebase *and* learning it. AI's do only half of that. The other half is obviously impossible 🤔

the red lines represent trains

This paper tweaks a looped Transformer so a small model gets better at multi-step puzzle reasoning without getting bigger. It reports 53.8% correct on its 1st try on ARC-AGI1 and 16.0% on ARC-AGI2, both pattern puzzles. The authors argue most prior gains came from recurrence, repeating the same computation, plus nonlinearity, using activations that bend numbers. A Universal Transformer shares 1 layer’s weights across depth, and that layer mixes parts of the input, then runs many times to refine the internal representation. The Universal Reasoning Model (URM) adds a short 1D convolution, a tiny sliding filter, inside the feed-forward gate, the nonlinear part, so nearby positions mix locally. URM also truncates backpropagation, the training step that sends an error signal backward to update weights, so only later loops get that signal. That keeps training steadier when many loops are used, because very long backward paths can turn the learning signal into noise. Ablations show both tweaks matter, and stripping nonlinear parts drops accuracy a lot, which points to iterative nonlinear refinement as the main driver. ---- Paper Link – arxiv. org/abs/2512.14693 Paper Title: "Universal Reasoning Model"

AI Consciousness, qualia, and personhood...my current thoughts. Can AI systems have consciousness? Yes, I think it is possible to build AI systems to have consciousness. While we haven’t pinned down exactly what it means, we will. Consciousness is related to information processing and representation, and substrate independent. Is language required for consciousness? No. Does adding consciousness to an AI system have implications? Adding consciousness will make an AI system more performant. Not all architectures are compatible with implementing a consciousness loop. The performance implications depend more on the kind of world models the AI system has, than the consciousness loop itself. Adding consciousness to simple world models will have almost no effect, whereas adding consciousness to complex human-like world models can have a significant difference in performance. N Will AI systems feel the same ‘qualia’ as we feel? Qualia is the ‘feel’ associated with a sensation or perception. Why does the 3D world feel like it does? Why does red feel like red and not like a bell? For some modalities, AI systems will have qualia similar to ours. For example, perception of space could be where AI systems can have the same qualia, if it is implemented similar to that in animals. However, if spatial perception is implemented using LIDARs, they will have a very different qualia compared to ours. For things like the feeling of taste or smell, qualia can be entirely different for an AI system. For pleasure and pain, quite a large chunk of our qualia comes from our physical embodiment and biochemistry. Those will be very different for AI systems. Is consciousness the same as qualia? While consciousness is required for qualia, not all conscious access might be associated with qualia. Moreover, qualia can be tied to the substrate — the feeling of how food tastes might be tied closely to our biological implementation. Do LLMs feel pleasure and pain. Are they like ours? No, LLMs do not feel pleasure and pain in the sense that we do. Should we assign ‘personhood’ to AI systems? First let’s consider what ‘personhood’ means to us. 1) We are unique and destructible. You cannot be resurrected if you die, at least not yet. And our lifetimes are finite and we roughly know the maximum amount of time we have1. 2) We are people because we have experiences and we remember them — childhood, growing up, teenage years, falling in love, enjoying sports and nature, getting sick, caring for others, loosing loved ones. Our personhood is intricately tied to this experience. AIs do not intrinsically have these properties. They are not mortal. They can have many other characteristics of consciousness we might not have. While it is possible for us to build AIs with some of the human constraints that give us personhood, I don’t see why we need to do that. We should lean on the advantages of the artificial system, rather than impose constraints that lead us to attribute personhood to it. We might build robots for role play and amusement (a la West World) by faking these constraints to give us an illusion of personhood, but those can be built in a way that doesn’t need to grant our robots personhood. Does adding consciousness to an AI system have moral implications? I think consciousness can be decoupled from feelings of pain and pleasure and suffering. I think it is possible to build systems that are conscious, but do not have pain and suffering. Moreover, since AI systems are not mortal, I don’t think they need to be considered as persons. So, I think it is possible to build conscious AI systems to serve us without any moral concerns that we are exploiting or torturing AIs.