schemes for approximating Solomonoff or AIXI look like at least exponential brute force search.
Well, yeah. Again—why would you expect anything else? Given that there exist problems which require that or worse for solution? How can a universal problem solver do any better?
Since AIXI is, by construction, the best possible intelligent agent, all work on AGI can, in a rather useless sense, be described as an approximation to AIXI.
Yes.
To the extent that such an attempt works (i.e. gets substantially further than past attempts at AGI), it will be because of new ideas not discovered by brute force search, not because it approximates AIXI.
No. Given how strange and different AIXI works, it can easily stimulate new ideas.
No. Given how strange and different AIXI works, it can easily stimulate new ideas.
The spin-off argument. Here’s a huge compendium of spinoffs of previous approaches to AGI. All very useful, but not AGI. I’m not expecting better from AIXI.
Hm, so let’s see; you started off mocking the impossibility and infeasibility of AIXI and any computable version:
I am not persuaded that the harder Bayesians have any more concrete answer. Solmonoff induction is uncomputable and seems to unnaturally favour short hypotheses involving Busy-Beaver-sized numbers. And any computable approximation to it looks to me like brute-forcing an NP-hard problem.
Then you admitted that actually every working solution can be seen as a form of SI/AIXI:
There might well be a theorem formalising that statement. There might also be one formalising the statement that every remotely optimal form of induction or decision-making is uncomputable. If that’s the way it is, well, that’s the way it is… Since AIXI is, by construction, the best possible intelligent agent, all work on AGI can, in a rather useless sense, be described as an approximation to AIXI
And now you’re down to arguing that it’ll be “very useful, but not AGI”.
I stand by the first quote. Every working solution can in a useless sense be seen as a form of SI/AIXI. The sense that a hot-air balloon can be seen as an approach to landing on the Moon.
And now you’re down to arguing that it’ll be “very useful, but not AGI”.
At the very most. Whether AIXI-like algorithms get into the next edition of Russell and Norvig, having proved of practical value, well, history will decide that, and I’m not interested in predicting it. I will predict that it won’t prove to be a viable approach to AGI.
Well, yeah. Again—why would you expect anything else? Given that there exist problems which require that or worse for solution? How can a universal problem solver do any better?
Yes.
No. Given how strange and different AIXI works, it can easily stimulate new ideas.
It’s more than I had before.
The spin-off argument. Here’s a huge compendium of spinoffs of previous approaches to AGI. All very useful, but not AGI. I’m not expecting better from AIXI.
Hm, so let’s see; you started off mocking the impossibility and infeasibility of AIXI and any computable version:
Then you admitted that actually every working solution can be seen as a form of SI/AIXI:
And now you’re down to arguing that it’ll be “very useful, but not AGI”.
Well, I guess I can settle for that.
I stand by the first quote. Every working solution can in a useless sense be seen as a form of SI/AIXI. The sense that a hot-air balloon can be seen as an approach to landing on the Moon.
At the very most. Whether AIXI-like algorithms get into the next edition of Russell and Norvig, having proved of practical value, well, history will decide that, and I’m not interested in predicting it. I will predict that it won’t prove to be a viable approach to AGI.
How can a hot air balloon even in theory be seen as that? Hot air has a specific limit, does it not—where its density equals the outside density?