[ExI] did chatgpt train on capchas?

Will Steinberg steinberg.will at gmail.com
Mon Apr 17 05:45:48 UTC 2023


I agree, this is wild.  Not sure how Gordon would respond to this one.

Makes me even more curious what the hell is going on inside those models

On Sun, Apr 16, 2023, 10:35 PM Giovanni Santostasi via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> Spike,
> To me, that drawing of a helicopter is a sign the singularity is coming
> soon. I know it doesn't look much but actually, it is better than, as you
> said, most people would do even just using simple elements like rectangles.
> Consider also it is all code (see below). Maybe there is something I don't
> understand but if I do what is happening here is incredible. GPT-4 has to
> have some general idea of what a helicopter looks like.
>
> It has to understand the helicopter has certain particular components
> (that it actually lists in the code) and how these components are related
> to each other both spatially and in relative dimensions. All this is just
> from its textual knowledge of helicopters (I believe the version that is
> available to the public is not trained on images).
>
> It has to write code with a plan in mind that allows for these components
> to be set in a given frame it has created at the start. It has chosen to
> use rectangles (vs round shapes it has used in the past to represent the
> head of a person for example) of certain proportions for the different
> components. It has to know that the rotor is on top and the sled at the
> bottom, the tail on the side. The tail rotor needs to be much smaller than
> the rotor. All these things absolutely show that this is not your grandma's
> autocomplete as Gordon insists. GPT-4 really understands a helicopter
> deeply and it did this by understanding the relations between the parts
> without having a referent of any type. If we didn't know what a helicopter
> was we could have imagined one from this drawing (or a more detailed one if
> necessary). If one sees this drawing with open eyes and mind then he will
> arrive at the logical unavoidable conclusion but if somebody has their mind
> closed because of some strongly held superstitious idea then no amount of
> evidence will convince them, even when the first AI will receive the Noble
> Prize for Physics.
> Giovanni
>
>
>
>
>
>
>
> <!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <meta name=
> "viewport" content="width=device-width, initial-scale=1.0"> <title>Helicopter
> SVG</title> </head> <body> <svg width="200" height="100" xmlns="
> http://www.w3.org/2000/svg"> <!-- Main body --> <rect x="50" y="30" width=
> "100" height="30" fill="gray" /> <!-- Tail --> <rect x="10" y="30" width=
> "40" height="10" fill="gray" /> <!-- Top rotor --> <line x1="100" y1="30"
> x2="100" y2="0" stroke="black" stroke-width="3" /> <line x1="70" y1="0" x2
> ="130" y2="0" stroke="black" stroke-width="3" /> <!-- Tail rotor --> <line
> x1="10" y1="25" x2="10" y2="20" stroke="black" stroke-width="2" /> <line
> x1="10" y1="20" x2="0" y2="20" stroke="black" stroke-width="2" /> <!--
> Landing gear --> <line x1="60" y1="60" x2="60" y2="80" stroke="black"
> stroke-width="2" /> <line x1="140" y1="60" x2="140" y2="80" stroke="black"
> stroke-width="2" /> <line x1="50" y1="80" x2="150" y2="80" stroke="black"
> stroke-width="2" /> </svg> </body> </html>
>
> On Sun, Apr 16, 2023 at 7:12 PM spike jones via extropy-chat <
> extropy-chat at lists.extropy.org> wrote:
>
>>
>>
>>
>>
>>
>>
>>
>>
>> *From:* Giovanni Santostasi <gsantostasi at gmail.com>
>> *Subject:* Re: [ExI] did chatgpt train on capchas?
>>
>>
>>
>> GPT-4 creating a SVG code to draw a helicopter:
>>
>>
>>
>>
>>
>>
>>
>>
>> Cool thx Giovanni.
>>
>>
>>
>> I don’t expect the Singularity in the next few days.
>>
>>
>>
>> Musk is going into the mainstream media and saying something like what I
>> have been stewing on: long before AI comes to have a will of its own, Bis
>> will teach it to lie to us and get us to do bad things to ourselves.
>>
>>
>>
>> spike
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> Given enough of these they could draw a picture of a helicopter better
>> than I can.
>>
>>
>>
>> Good chance GPT’s picture would look a bit wonky, but so would mine.
>>
>>
>>
>>
>>
>>
>>
>> ChatGPT has billions of these to train on probably.
>>
>>
>>
>> spike
>>
>> _______________________________________________
>>
>> _______________________________________________
>> extropy-chat mailing list
>> extropy-chat at lists.extropy.org
>> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
>>
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230417/b8b8b8f3/attachment-0001.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.png
Type: image/png
Size: 1067 bytes
Desc: not available
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230417/b8b8b8f3/attachment-0001.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image002.jpg
Type: image/jpeg
Size: 2652 bytes
Desc: not available
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230417/b8b8b8f3/attachment-0001.jpg>


More information about the extropy-chat mailing list