The original Turing Test

[John Clark]

In Alan Turing's original paper in which he propose the thing that we now call the Turing Test he gave this example of the sort of thing we might ask a machine that claims to be intelligent and the answer we might expect to receive:

Question: Please write me a sonnet on the subject of the Forth Bridge.

Answer: Count me out on this one. I never could write poetry.

So I decided to give the the same order to ChatGPT and less than five seconds it did so. 

Question: Please write me a sonnet on the subject of the Forth Bridge.

Answer:  Amidst the Scottish skies and bracing breeze,
A structure stands with towering grace and might,
A symbol of Victorian expertise,
A masterpiece of steel, a wondrous sight.

The Forth Bridge, a marvel of its time,
 A cantilevered railway, bold and strong,
Its spans reach far, a feat of man's design,
Defying nature's forces all day long.

For over a century it has stood,
A steadfast sentinel of the North Sea,
Its steel beams gleaming bright in sun and flood,
A tribute to the skill of industry.

Oh, Forth Bridge, a wonder of the age,
A testament to man's unyielding rage.

I'm the world's worst poet so I'm not a good judge of the above's artistic value but I'm certain it's better than what Alan Turing would have expected a machine that claims to be intelligent would say. Now if it could just do arithmetic as well as a two dollar calculator!

John K Clark    See what's on my new list at  Extropolis

56w

[Telmo Menezes]

It has long been observed in the field of AI that often things that look hard end up being the easy ones, and things that look easy turn out to be surprisingly hard.

We now have amazing statistical AI (current stuff) and amazing symbolic AI (chess engines, etc), but we do not really know how to combine the two. I tend to connect this to the ideas of Daniel Kahneman ("Thinking, fast and slow"). I think that statistical (roughly) corresponds to the fast mode of thinking and symbolic to the slow mode. Or maybe this is a bit simplistic and silly, but the important thing is that we humans can combine and switch between the two modes of cognition, and I think this is crucial for AGI.

Telmo

John K Clark    See what's on my new list at  Extropolis

56w

--

You received this message because you are subscribed to the Google Groups "Everything List" group.

To unsubscribe from this group and stop receiving emails from it, send an email to everything-li...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/everything-list/CAJPayv0s3AEXh7O%3DcUGz%3D8cMGbnAZ6xYhFCaHhYyyvnFtiRo3g%40mail.gmail.com.

[Jason Resch]

On Wed, Mar 1, 2023, 10:06 AM Telmo Menezes <te...@telmomenezes.net> wrote:

Am Mi, 1. Mär 2023, um 15:50, schrieb John Clark:

In Alan Turing's original paper in which he propose the thing that we now call the Turing Test he gave this example of the sort of thing we might ask a machine that claims to be intelligent and the answer we might expect to receive:

Question: Please write me a sonnet on the subject of the Forth Bridge.

Answer: Count me out on this one. I never could write poetry.

So I decided to give the the same order to ChatGPT and less than five seconds it did so. 

Question: Please write me a sonnet on the subject of the Forth Bridge.

Answer:  Amidst the Scottish skies and bracing breeze,
A structure stands with towering grace and might,
A symbol of Victorian expertise,
A masterpiece of steel, a wondrous sight.

The Forth Bridge, a marvel of its time,
 A cantilevered railway, bold and strong,
Its spans reach far, a feat of man's design,
Defying nature's forces all day long.

For over a century it has stood,
A steadfast sentinel of the North Sea,
Its steel beams gleaming bright in sun and flood,
A tribute to the skill of industry.

Oh, Forth Bridge, a wonder of the age,
A testament to man's unyielding rage.

I'm the world's worst poet so I'm not a good judge of the above's artistic value but I'm certain it's better than what Alan Turing would have expected a machine that claims to be intelligent would say. Now if it could just do arithmetic as well as a two dollar calculator!

It has long been observed in the field of AI that often things that look hard end up being the easy ones, and things that look easy turn out to be surprisingly hard.

We now have amazing statistical AI (current stuff) and amazing symbolic AI (chess engines, etc), but we do not really know how to combine the two. I tend to connect this to the ideas of Daniel Kahneman ("Thinking, fast and slow"). I think that statistical (roughly) corresponds to the fast mode of thinking and symbolic to the slow mode. Or maybe this is a bit simplistic and silly, but the important thing is that we humans can combine and switch between the two modes of cognition, and I think this is crucial for AGI.

I agree there's some relation.

Everything ChatGPT does is in a single evaluation of its neural network. It is like being asked to look at a chess board and immediately name the first best move that comes to your mind instinctively, rather than trying to progress through the game tree iteratively and see where it goes.

This is also why it fails at multiplication, which is what computers scientists call an Order-N-squared problem, meaning the number of steps an algorithm must takes to perform multiplication grows roughly by N^2 as the input size of the multiplicands grows by N.

Since ChatGPT always uses a constant number of steps in it's processing, there is some size N where it necessarily fails at multiplication. This is true for any machine/algorithm that use a constant number of steps.

The solution for ChatGPT, as well as for humans, is to break the problem down into manageable steps and work on them piece meal until the problem is solved. This is why mathematicians need chalkboards to do what they do, multiplying big numbers or solving proofs, often requires many steps and can't be solved by intuition alone.

Google discovered this with their alpha zero chess/go AI: It's single network evaluation of the best move had an ELO score of around 3000, still better than the best humans, but not by much. But by letting it iterate through different top moves and see which course of action faired best overtime, it increased it's level of play to 5000 ELO, well into super human territory.

For reference a beginner chess player has an ELO of 1000, a decent amateur around 1200, a chess master 1800, grand masters start around 2400, while the best human players in history have been around 2800.

So I think we need to make a "meta ChatGPT wrapper" that asks ChatGPT how to break down a problem into smaller steps, query ChatGPT multiple times to solve each step and also double check it got the right result along each step in the process, if it second guesses itself, ask it to break down that intermediate step, and so on recursively, until the problem is solved.

I think this replicates the thought processes for a wide range of human intellectual activity and I believe it would widen the class of problems ChatGPT could effectively solve.

Jason

[spudb...@aol.com]

Turing would have been enthused to the Max. Now could he have done something to test it out?? What would Turing 1950 have concluded?