Simulated Intelligence

Apple's recent paper has kicked off a kind of Rorschach discussion about whether LLMs actually reason or just pattern match.
For me, I recognise this when asking Claude Code to think, think harder or ULTRATHINK. If you use the sledgehammer on a nut-sized problem, you'll have a bad time.
But more than that, my favourite theme from the piece was the sense of "simulation".
Everyone knows what makes a simulator different from real life. It's enough like reality that it's useful, and it's useful because it lets you see what happens without cost or risk of failure.
However these tools work, they create truly useful simulations of intelligence.
What's the difference between a simulation of intelligence and real intelligence? In some ways EVERYTHING. In other ways, NOTHING.
Getting the most from these tools maybe depends on your ability to keep that in your own context window the whole time.