> The architectures of these models are a plenty good scientific basis for this statement.

That wouldn't be full-on science, that's just theoretical. You need to test your predictions too!

--

Here's some 'fun' scientific problems to look at.

* Say I ask Claude Opus 4.5 to add 1236 5413 8221 + 9154 2121 9117 . It will successfully do so. Can you explain each of the steps sufficiently that I can recreate this behavior in my own program in C or Python (without needing the full model)?

* Please explain the exact wiring Claude has for the word "you", take into account: English, Latin, Flemish (a dialect of Dutch), and Japanese. No need to go full-bore, just take a few sentences and try to interpret.

* Apply Ethology to one or two Claudes chatting. Remember that Anthropomorphism implies Anthropocentrism, and NOW try to avoid it! How do you even begin to write up the objective findings?

* Provide a good-enough-for-a-weekend-project operational definition for 'Consciousness', 'Qualia', 'Emotions' that you can actually do science on. (Sometimes surprisingly doable if you cheat a bit, but harder than it looks, because cheating often means unique definitions)

* Compute an 'Emotion vector' for: 1 word. 1 sentence. 1 paragraph. 1 'turn' in a chat conversation. [this one is almost possible. ALMOST.]