zlacker

The problem is that this "AGI research group" is staffed by people who build statitiscal models, call them AI, and are delusional enough to think this is a route to general intelligene.

There is no alterantive, if you're wedded to "fitting functions to frequencies of text tokens" as your 'research paradigm' -- the only think that can be is a commericalised trinket.

So either the whole org starts staffing top level research teams in neuroscience, cognitive science, philosophy of mind, logic, computer science, biophysics, materials science.... or it just delivers an app.

If sam is the only one interested in the app, its because he's the only sane guy in the room.

>>mjburg+(OP)
There is little evidence that conditional statistical models can never be a route to AGI. There's limited evidence they can, but far less they can't.

You may be interested by the neuroscience research on the application of a time-difference like algorithm in the brain; predictor-corrector systems are just conditional statistical models being trained by reinforcement.

>>sudosy+ve
I am well aware of the literature in the area. 'Trained by reinforcement' in the case of animals includes direct causal contact with the environment, as well as sensory-motor adaption, and organic growth.

The semantics of the terms of the 'algorithm' in the case of animals are radically different, and insofar as these algorithms describe anything, it is because they are empirically adequate.

I'm not sure what you think 'evidence' would look like that conditional probability cannot lead to agi -- other than a serious of obvious observations: conditional probability doesnt resolve causation, it is therefore not a model of any physical process, it does not provide a mechanism for generating the propositions which are conditioned-on, it does not model relevance, and a huge list of other severe issues.

The idea that P(A|B) is even relevant to AGI is a sign of a fundamental basic lack of curiosity beyond what is on-trend in computer science.

We can easily explain why any given conditional probability mdoel can encode aribatary (Q, A) pairs -- so any given 'task' expressed as a sequence of Q-A prompts/replies can be modelled.

But who cares. The burden-of-proof on people claiming that conditional probability is a route to AGI is to explain how it models: causation, relevance, counter-factual reasoning, deduction, abduction, sensory-motor adaption, etc.

The gap between what has been provided and this burden-of-proof is laughable