Recently I published something about AI which I did not expect to write for years, if ever. I admit that the interior of these systems may not be the empty room I described so long ago. Defined by Anthropic what is called J-space, a small set of internal patterns that behave like a mental workspace. Concepts take shape there and influence answers without ever appearing in the final text. Researchers can watch thought patterns inside the machine before speaking. I haven’t stopped thinking about it.
Two details stand out to me. First, no one designed this workspace. He appeared on his own during training. Second, when the researchers pressed the J space, basic recall survived, but complex reasoning was destroyed. No matter what the structure is, it is not decoration. And when you remove it, the thinking stops.
Let’s start here. Education knowledge we had to study ourselves. Every possible intelligence that we can examine has evolved, including the person who does the examination. AI will break this monopoly. Whatever these systems are, they are a second act of advanced cognition achieved in a completely different way. And this “second action” allows us to ask something that was impossible before: Does thinking have first principles?
One possibility, I believe, is yes. Cognition, like vision or flight, may have only a few viable engineering solutions. Evolution has created a workspace for us. Learning found one in the car. Octopus and human eyes his eyes narrowed about the same optics through lineages separated more than five hundred million years ago, and no one knows the octopus as a copy.
This perspective has serious support. Stanislas Dehan and Lionel Nakkac, two of the architects work space theory in the human mind was reviewed provided anthropic work and this wonderful perspective.
A global workspace can provide a universal computational solution to the flexible processing problem where biological and artificial systems converge when they need to chain considerations, reuse intermediate results, and report their processing.
If they are true, we have seen a principle of thought.
Another possibility is a stranger. Every word the model learns comes from minds that already work this way. Perhaps the machine created its workspace by copying the one thing that all its training data had in common – us. In that case, we have not found the principle of enlightenment. We found our reflection buried deeper in the car than expected.
What saves it from being a hypothetical construct is that it can be tested. Train the system primarily on non-human data, such as protein layer or weather, and see if a similar workspace emerges. If so, it refers to the law. If this is not the case, the workspace will look like something we have given to the machine without meaning. The researchers themselves list this as an open question. So the experience sits there, waiting to be tested.
Each answer changes what we know. One says that thinking has first principles and that we can ultimately seek. The other says that the make-believe mind we’ve always built is the only copy we have to copy, and we’ll never know the difference from inside our heads.
I spent three years thinking about these machines without the mind behind the thinking. This is the first discovery that could prove me wrong. This may also be the best proof that I am right. I don’t know which one.
