I think that this should be in a framework that takes into account granularity in some sense. Like I assume you are thinking about the kolmogorov complexity of simulating a system observationally similar to
Case 1: This is either A: A generic, realistic looking adult brain (hard to estimate the complexity), vs B: the brain of an individual person (~amount of synapses).
When you say claude opus is more intelligent than haiku, in case B it would definitely have more complexity, but in case A:
---If opus and Haiku were trained on the same dataset they would have almost the same complexity except for the num_parameters in the code. (Opus would have 1-2 bits more)
---More interestingly, it could be that Opus seems to have more geenral intelligence, rather than just knowledge because the architecture is more expressive and can learn an underlying algorithm that is a little bigger, but if you simulated Opus and Haiku with a more advanced architecture, maybe both would have the same “raw intelligence”. This is related to the texture vs shape bias in image classifiers. Models above 1B start recognizing stuff by shape rather than texture, which seems more like a simple algorithmic improvement rather than something that fundamentlaly required a higher parameter count.
Case 2: your subjective experience (could be compiled into a list of brain activations and sensory data)
Case 3: a generic human baby brain.
Case 4: Step 1: take laws of physics + a PRNG, simulate a universe, presumably it will have some intelligences. Step 2: build an “intelligence detector”, something that can detect human-like civilizations, e.g by finding complex radio emissions, then seek the brains somehow. This likely fits in less than 1MB.
I think that this should be in a framework that takes into account granularity in some sense. Like I assume you are thinking about the kolmogorov complexity of simulating a system observationally similar to
Case 1: This is either A: A generic, realistic looking adult brain (hard to estimate the complexity), vs B: the brain of an individual person (~amount of synapses).
When you say claude opus is more intelligent than haiku, in case B it would definitely have more complexity, but in case A:
---If opus and Haiku were trained on the same dataset they would have almost the same complexity except for the num_parameters in the code. (Opus would have 1-2 bits more)
---More interestingly, it could be that Opus seems to have more geenral intelligence, rather than just knowledge because the architecture is more expressive and can learn an underlying algorithm that is a little bigger, but if you simulated Opus and Haiku with a more advanced architecture, maybe both would have the same “raw intelligence”. This is related to the texture vs shape bias in image classifiers. Models above 1B start recognizing stuff by shape rather than texture, which seems more like a simple algorithmic improvement rather than something that fundamentlaly required a higher parameter count.
Case 2: your subjective experience (could be compiled into a list of brain activations and sensory data)
Case 3: a generic human baby brain.
Case 4: Step 1: take laws of physics + a PRNG, simulate a universe, presumably it will have some intelligences. Step 2: build an “intelligence detector”, something that can detect human-like civilizations, e.g by finding complex radio emissions, then seek the brains somehow. This likely fits in less than 1MB.