I was just turned on to this 1956 paper by W Ross Ashby on the subject of whether it is possible for someone to generate a machine more intelligent than themselves and I think you will like it a great deal
The Turing 1950 paper anti-recommendation resonated with something I have just written on how (quoting you) "everyone projects their own pet theories and predilections onto it, often not even noticing that what they are saying is in direct contradiction with what Turing was writing."
I'm becoming converted to the "flower arranging" school* of philosophical writing. What a lovely bouquet!
* This school does not exist afaik, but in the same way a koan can arrange words that capture or gesture at a unique meaning in the very composition, why not express one's philosophical perspective by a similar work of arrangement?
Great picks, the anti recommendation made me chuckle.
What I didn't quite understand was your remark on SLT now being expedient and workable because formalized in LEAN. In the referenced paper I read that a "human-AI collaborative workflow" generated 30k lines of LEAN code which executed without sorries. Surely this does not make applying SLT to practical problems any easier.
That part was a bit tongue-in-cheek. Basically, all the empirical process theory stuff is beautiful mathematics, but it is also the most mechanical part of the theory (which, as Ben Recht and I argued in our recent paper*, is often unnecessary).
Nice list of papers! I don’t quite understand your criticism of vagueness in Turing’s paper, though. I find the imitation game refreshingly precise -- Turing replaces an ill-posed metaphysical question with an operational comparison, and imposes the same evidential standard on humans and machines. To me, this is the most insightful contribution to the whole intelligence discussion.
I'm just a simple electrical engineer from a backwoods Midwest university, and to me Turing's comparison is not operational, it's circular. Essentially, it says something like "system S is actually thinking if, based on interacting with it, we think that it's thinking." An operational criterion should be formulated solely in an object language, but Turing's criterion sneaks in a metalinguistic construct -- using intelligence to adjudicate intelligence. What happens if we cut out the human evaluator altogether? If we grant that the ability to conduct the Turing test and to evaluate its results a form of intelligence, then presumably it can itself be simulated by a machine, so we will have machines evaluating machines evaluating machines etc. It's the same infinite regress that arises when we attempt to justify induction using meta-induction, which in turn requires meta-meta-induction to justify it, etc.
You’re right that he’s not offering a non-circular definition of intelligence. What he’s really doing is offering an epistemic handle—a way to know that a machine is intelligent, without making any assumptions about its internal structure.
There is no vicious circle here because it is a way to go from a stable knowledge base (we know our ordinary conversation partners are intelligent) and to generalize beyond that (to machines).
As a pragmatic shortcut, it's perfectly fine (but then we have to acknowledge that ELIZA had passed the Turing test with flying colors). But it does not provide an operational criterion because there cannot be one (just like there can be no rock-solid justification for induction).
I was just turned on to this 1956 paper by W Ross Ashby on the subject of whether it is possible for someone to generate a machine more intelligent than themselves and I think you will like it a great deal
https://gwern.net/doc/ai/1956-ashby.pdf
Ah yes, the intelligence amplifier paper! I do indeed like it a lot.
Great post, as usual.
The Turing 1950 paper anti-recommendation resonated with something I have just written on how (quoting you) "everyone projects their own pet theories and predilections onto it, often not even noticing that what they are saying is in direct contradiction with what Turing was writing."
https://mariofigueiredo.substack.com/p/what-turing-didnt-say
I'm becoming converted to the "flower arranging" school* of philosophical writing. What a lovely bouquet!
* This school does not exist afaik, but in the same way a koan can arrange words that capture or gesture at a unique meaning in the very composition, why not express one's philosophical perspective by a similar work of arrangement?
Great picks, the anti recommendation made me chuckle.
What I didn't quite understand was your remark on SLT now being expedient and workable because formalized in LEAN. In the referenced paper I read that a "human-AI collaborative workflow" generated 30k lines of LEAN code which executed without sorries. Surely this does not make applying SLT to practical problems any easier.
That part was a bit tongue-in-cheek. Basically, all the empirical process theory stuff is beautiful mathematics, but it is also the most mechanical part of the theory (which, as Ben Recht and I argued in our recent paper*, is often unnecessary).
*https://arxiv.org/abs/2604.19560
Nice list of papers! I don’t quite understand your criticism of vagueness in Turing’s paper, though. I find the imitation game refreshingly precise -- Turing replaces an ill-posed metaphysical question with an operational comparison, and imposes the same evidential standard on humans and machines. To me, this is the most insightful contribution to the whole intelligence discussion.
I'm just a simple electrical engineer from a backwoods Midwest university, and to me Turing's comparison is not operational, it's circular. Essentially, it says something like "system S is actually thinking if, based on interacting with it, we think that it's thinking." An operational criterion should be formulated solely in an object language, but Turing's criterion sneaks in a metalinguistic construct -- using intelligence to adjudicate intelligence. What happens if we cut out the human evaluator altogether? If we grant that the ability to conduct the Turing test and to evaluate its results a form of intelligence, then presumably it can itself be simulated by a machine, so we will have machines evaluating machines evaluating machines etc. It's the same infinite regress that arises when we attempt to justify induction using meta-induction, which in turn requires meta-meta-induction to justify it, etc.
You’re right that he’s not offering a non-circular definition of intelligence. What he’s really doing is offering an epistemic handle—a way to know that a machine is intelligent, without making any assumptions about its internal structure.
There is no vicious circle here because it is a way to go from a stable knowledge base (we know our ordinary conversation partners are intelligent) and to generalize beyond that (to machines).
As a pragmatic shortcut, it's perfectly fine (but then we have to acknowledge that ELIZA had passed the Turing test with flying colors). But it does not provide an operational criterion because there cannot be one (just like there can be no rock-solid justification for induction).