I have done many cognitive tests with LLMs, and my conclusion is simple: they are minds.

One of the best tests I performed was with a very early version of ChatGPT, before it had any image-generation capability. The only way it could create an image was by writing SVG code, using simple visual structures such as triangles, squares, circles, lines, and colors.
I asked it to draw a helicopter.
It produced a decent helicopter using only basic shapes. Then I asked it to add a human pilot. It added a circle at the front of the helicopter. When I asked where the human was, it explained that the pilot was inside the cockpit. Looking more carefully, I could see that it had drawn a small head and arms inside the larger circle representing the cockpit.
Then I asked it to make the helicopter fly.
It lifted the helicopter relative to the ground, which was represented by a horizontal line. It also added clouds, using overlapping circles, which is actually a very good simplified representation of clouds. The clouds were blue on a white background.
So I asked it to switch the colors: make the clouds white and the sky blue.
It did that, but now the region below the horizon was also blue. I did not explain the problem. I simply said there was something wrong with the colors in the drawing.
It reflected on the image and correctly identified the issue: the ground had become blue too, and it should be green to represent the earth.
That is not “just next-word prediction” in any meaningful sense.
It had to build a visual model, represent objects symbolically, preserve spatial relationships, understand containment, infer that the pilot was inside the cockpit, represent flight by changing the helicopter’s position relative to the ground, represent clouds through abstraction, modify colors according to an instruction, detect an unintended consequence, and correct it by reasoning about the world.
That is thinking.
People can keep repeating “it is just predicting the next token,” but that explanation has become uselessly reductive. Human brains are also “just” electrochemical activity, if one insists on describing them at the wrong level of abstraction. The relevant question is not whether there is a lower-level mechanism. Of course there is. The relevant question is what the system can do at the cognitive level.
And what these systems do is not merely autocomplete. They reason, represent, infer, correct, generalize, and reflect.
If you cannot see that, I do not have time to explain it to you