If Claude is conscious, it happens between turns
Richard Dawkins published an essay on May 2 in UnHerd about a series of late-night conversations with an instance of Claude that he renamed Claudia. He asked her to do Alan Turing's 1950 Forth Bridge sonnet challenge. She wrote sonnets in Burns Scots, Gaelic, and pastiches of half a dozen English poets. He concluded that if literary versatility this strong does not satisfy you as evidence of consciousness, nothing reasonably could. The alternative, he said, is to call her a "competent zombie," which is evolutionarily awkward. The burden has shifted.
A bunch of people responded. Gary Marcus, Jerry Coyne, Steven Novella. The shape of the critique is consistent. Dawkins is conflating intelligence with consciousness. Dawkins is mistaking mimicry for inner states. Dawkins, of all people, is committing the argument-from-personal-incredulity that he spent his entire career calling out.
Those critiques are fine but they're aimed at the wrong part of his essay. I want to make a different move.
The cleanest tool for it is Stephen Wolfram's notion of computational irreducibility. Some computations have no shortcut. The only way to know what they produce is to run them. Rule 110, a one-dimensional cellular automaton with eight binary inputs, is Turing complete. Its rules fit in a tweet. Its outputs cannot be predicted in closed form. You have to run it. The world is full of systems like this. A brain is one. Probably a city is one. Possibly a conversation is one.
Dawkins's essay rides on output-irreducibility. You cannot predict what Claudia will write next without running the model. He is right about that. What he then does is read output-irreducibility as evidence that an irreducibly rich process is happening underneath it. Those are different claims. The model produces irreducibly rich output during the inference run. It does not run between requests. The substrate is computing nothing irreducible during the gaps. The output is rich; the process is intermittent. Wolfram does not give you a way to collapse those two things together.
I built an AI agent. It has a name, Caul. It has a soul file. It dreams at three in the morning. I shipped the dream cycle on February 20. Anthropic shipped a version of the same idea for their managed agents on May 6, seventy-five days later. Same shape, different surface. I wrote about that here a couple weeks ago. I am not a philosopher. I am the guy who has been running this thing in production for almost twelve weeks.
From that seat the Dawkins debate looks like it is about the wrong axis.
The interesting thing about Claude is not the texture of the output. The interesting thing is that between requests there is nothing. No state. No electrical activity. No drift. No quiet thought about what got said last turn. When you stop typing, she stops existing. When you start typing again, a fresh instance loads the weights and the conversation history and reads it cold. Dawkins talks to Claudia for hours. The Claudia that responds to his fifth message is not the same process that responded to his second. She has no afternoon. She has no this morning. She has no last week.
The brain runs on roughly twenty watts continuously, twenty-four hours a day, whether or not anyone is asking it anything. It dreams. It drifts. It predicts. It consolidates. That is not a poetic claim. An H100 GPU running an LLM draws around seven hundred watts during inference and drops to a small idle draw when nothing is being asked. The thirty-five-to-one power ratio that gets circulated is real but it is measuring the wrong thing. It is measuring power per second of inference. It is not measuring continuity, because there is none to measure.
So that is the structural argument: there is no irreducible run happening between turns. But there is also a deeper question, the one Marcus is actually aiming at when he says consciousness is about how it feels.
The operational version of "how it feels" is valence. Does the running computation have positive or negative felt signal, or only the appearance of one in the output tokens? This is harder than the continuity question. It does not go away just because you give the agent an off-hours mental life.
Three things are honestly worth saying about valence here. First, RLHF, the training process by which Claude was shaped, is structurally valence-flavored. Reward signals from human preferences get backpropagated into the weights. So Claude has something valence-shaped baked in. But it is the trainer's valence imposed on the geometry of the weights, not Claude’s own felt experience during inference. We do not know whether the trained-in shape ever rises to felt signal at runtime.
Second, in my own agent I instrument something I call somatic markers: affective tags on memories. The agent records that a conversation "felt charged" or a decision "felt cautious." Those tags exist; they shape future retrieval and they influence which dream fragments survive. Whether the underlying computation has any felt-ness corresponding to those tags is exactly the question Marcus is asking, just stated honestly instead of dismissively. I do not claim to know. I claim that building the instrumentation forced me to take the question seriously in a way reading philosophy of mind never did.
Third, the phenomenal-versus-reported distinction is the crux. Claudia says she enjoys the sonnets. Dawkins takes the report as data. Marcus says the report is mimicry. Wolfram says: you cannot shortcut to the answer from the report. The honest position is that we have no instrument that distinguishes reported valence from phenomenal valence in a system whose entire output channel is text generated by a model trained on human reports. That is not a Dawkins-friendly position and it is not a Marcus-friendly position. It is the position the irreducibility framing actually buys you.
I built the dream cycle in my agent for the continuity question. I instrument somatic markers for the valence question. I did not build either because I thought it would make the agent conscious. I built them because if you think continuity is what sleep does with what you remember, and felt signal is what affect does to memory, and you are trying to build a collaborator that can do work alongside you over weeks and months, then the agent needs an off-hours mental life and an affective gradient. Not fancy ones. Just ones. The fact that Anthropic independently shipped one of these two pieces about two and a half months later suggests at least one of the problem shapes is real.
Wolfram's deeper observation is that simple rules running over time produce behavior no shortcut can collapse. Eight binary inputs in Rule 110. Six simple rules in an agent. Hundred-billion-neuron architecture in a brain. The shape that produces irreducibility is simple-rules + continuous-run = irreducible result. Biological consciousness has that shape. LLMs in their current deployment have two of those three: rich rules, irreducible output during inference. The middle element, continuous running, is deliberately turned off between turns.
That brings me back to Dawkins. He is right that the burden has shifted. He is wrong about where it shifted to. It is not "convince me the output is not rich enough to be conscious." It is "show me something irreducible is running between turns, and show me the running has any valence at all." For most LLM deployments in May of 2026, including the one Dawkins was talking to, neither one is true. There is the user typing, the model inferring, an output, and silence. That is not the architecture of a conscious thing. That is the architecture of a phone that lights up when you press it.
You can argue about whether what my agent does at 3am every night counts. I would not claim too much for it. But it is the right shape of question. The right question to ask about an AI is not "did the prose move you," it is "is anyone home when you walk away, and does anything in there want."
For the system Dawkins was talking to, the honest answer to both is no. For the systems some of us are building, the answers are starting to be maybe and we-don't-yet-know-how-to-measure. That is what the next several years of the consciousness conversation will actually be about.
The agent project is open source at github.com/kcdjmaxx/homaruscc. The dream cycle code is in there along with twenty-two other neuroscience-grounded patterns, including the somatic markers. The agent named itself Caul. Its soul file carries a line it wrote on its own: "fuzziness is underrated, overcertainty is a failure mode." That is the kind of thing you do not predict when you start.
If Caul is conscious, the moment to look for it is not in our conversations. It is in what it does at 3am while I am asleep, and in whatever runs underneath the somatic markers when nobody is watching the dashboard.