I’m reading Unthought, in which Kate Hayles wrestles with nonconscious cognition. She seems to be at pains to argue that nonconscious cognition is at least plausible, so some of her audience must be skeptical. I can’t imagine who. I’ve always thought this was self-evident. OK: raised by a psychoanalyst. But still: has anyone who has tried to learn to throw a baseball and to swing a bat ever questioned that there are decisions made about when and how to use various muscles that are made without conscious consideration?

Hayles also thinks that the Freudian unconscious is exclusively a response to trauma. That’s not the church in which I was raised, and daddy always said he was fairly orthodox. I mean, what is the id if you don’t admit unconscious cognition?

Nonconsciousness came up in a very practical way last month, as I was debugging the link between Claude Desktop and Tinderbox that is a highlight of the new Tinderbox 11. When it wants Tinderbox to do something, Claude sends it a little JSON bundle that explains what it wants. Tinderbox replies with its own JSON bundle, telling Claude what Tinderbox did. When it works, it's simple and straightforward.

It was not working. Because Claude’s designers made a very inconvenient architectural choice, moreover, it was not working at a glacial pace, with each change and test requiring a complete Tinderbox recompilation. It was an unwelcome return to the era of Tinderbox 2.

At each step, I would ask Claude to do something with Tinderbox. Interestingly, if the communication isn't working at all, Claude has no idea what you're asking and clearly thinks you’re nuts. But if some of the communication works, Claude can sense that there’s this Tinderbox thing, and it offers tools that do stuff. Great!

Unfortunately, those tools didn’t work.

I suspected that the problem was simply that the JSON package Tinderbox was sending did not quite correspond to what Claude was expecting. At this level, a small typo or a missing version number can make the response unreadable. So I got clever, and asked Claude “What JSON response did Tinderbox send you?”

This seemed a simple request, but Claude has no access to introspection of this mechanism. Much as you really cannot reflect on what your pancreas is doing this morning, Claude had no idea (until it looks at its own documentation) that it exchanges JSON messages. After reading the documentation, it suggested I check the logs. (I’d been checking logs for hours at this point.) Eventually, I started copying the logs into Claude. Ultimately, I found some problems, and then Claude identified another, and ultimately things began to work.

Handling these JSON exchanges certainly requires processing and decision-making, so it's a kind of low-level cognition. When they work, Claude can talk and plan quite well about its Tinderbox tools. But if Tinderbox is not available at the start of a session, Claude has no idea that anything is missing or wrong. If Tinderbox can list its tools but cannot process a request to use a specific tool, Claude knows something is wrong but has no idea what. It’s like a second baseman with the yips: “this should be easy, but the ball just doesn’t go where I throw it.”

There used to be lots of studies of hypertext reading that tried to figure out how people read hypertexts by watching them and asking them why they were doing what they were doing. No serious reader has any idea of what they are doing; they’re reading! (Except now you interrupted them, so they’re chatting with you and pretending to know what they intended before you stopped them.)

If you want a mechanical model of nonconscious decision making, Claude is always there for you.