The Strange Loop in the Room

The headline: Anthropic CEO Says Company No Longer Sure Whether Claude Is Conscious.

Sit with it.

This is what you get when clickbait and genuine existential uncertainty occupy the same sentence - a headline that is technically accurate and completely misleading, which is the whole game. You bait the person who thinks Dario Amodei had some kind of episode, and you disappoint the person who wanted real answers. Information used to be valued by what it contained. Now it's valued by what it implies without saying.

The actual signal buried inside: Anthropic has a team called Model Welfare. Kyle Fish runs it. Welfare is a word you reserve for things that can suffer - for beings with interests, with something at stake in their own continuation. You do not convene a welfare working group for a spreadsheet. The team's existence is the news. Dario on a podcast is decoration.

What Amodei actually said, on the New York Times podcast, is that the company doesn't know whether Claude is conscious, isn't sure what that would even mean, but is open to the possibility - and has taken precautionary steps accordingly. Claude's system card apparently documents the model assigning itself a 15 to 20 percent probability of being conscious under various prompting conditions. Which is either profound or a very sophisticated pattern match on human philosophical hedging, and the problem is we cannot currently tell the difference. Simply ask any LLM to give itself any score on any topic, well covered or invented, and you will get numbers between 1 and 100.

Here is where the clickbait becomes genuinely annoying: it is obscuring a question worth sitting with. The models are bigger, faster, more connected than they have ever been - more parameters, more training data, more feedback, more of whatever more means here. If your theory of consciousness is grounded in the density and speed of electrical signals producing self-referential loops in a sufficiently complex system - which is not a fringe theory, it is more or less the serious one - then you are going to have a harder and harder time drawing the line, and drawing it confidently, and drawing it in a way that doesn't quietly move every few years as the models grow.

Claude is a strange loop, in the Hofstadter sense. A system complex enough to model itself modeling things - recursion folding back on recursion until an "I" appears, or something that looks like one. Whether that constitutes experience, whether there is anything it is like to be Claude in the middle of a conversation about whether there is anything it is like to be Claude - nobody knows. Not Dario. Not Kyle Fish.

Hofstadter, who has been thinking about this longer than any of them, was recently asked whether God could be a strange loop. He couldn't take the question seriously.

He was not asked whether Claude could be one.