[FRIAM] This is scary, and yet very cool...Ai neural networks making pictures, look really good
Russell Standish
lists at hpcoders.com.au
Sun Apr 10 21:12:54 EDT 2022
On Fri, Apr 08, 2022 at 08:19:16AM -0600, cody dooderson wrote:
> That soundscape thing is nice. What a nifty idea.
> Speaking of AI, has anyone used GitHub copilot. It is AI for writing code. It
> is spectacular. It writes decent code with very few prompts. It makes a few
> mistakes but don't we all.
> I haven't been able to find a good writeup on how it works. Does anyone know
> about it? Does it run locally? Could it create itself?
Yes - I've been reading some of the articles, and listened to some
talks presented at NVidia's GTC conference.
Copilot is a model based on GPT3, which is a general language parsing
and generation model. Basically, it takes input tokens, and outputs
most like response tokens based on that. It is a recurrent neural
network with a few 100 billion parameters (ie essentially synaptic
weights), with a complexity approaching that of a human brain (human
brains contain around a trillion synapses). GPT3 has been pretrained
on corpora sourced from the world-wide web, and interestingly it has
been found that the pre-trained model can be fairly rapidly retrained
on different tasks. In the case of Copilot, it has been trained on the
contents of Github, which is the world largest open source
repository. This has caused concern amongst copy left activists, as
quite a bit of GPL licensed code was involved, and this is arguably
not a case of "fair use" of copyright code.
No - it does not run locally. It requires a massive, massive cluster
of NVidia GPUs to run training, and the inference part requires a not
quite so massive cluster of GPUs as well. Hence the need for it to run
in the cloud.
>
> On Wed, Apr 6, 2022, 1:24 PM Marcus Daniels <marcus at snoutfarm.com> wrote:
>
>
> Change the sentence to be “Put Ukrainian soldier nearby a civilian on
> street in Bucha, shooting her.”
>
>
>
> From: Friam <friam-bounces at redfish.com> On Behalf Of Gillian Densmore
> Sent: Wednesday, April 6, 2022 11:09 AM
> To: The Friday Morning Applied Complexity Coffee Group <friam at redfish.com>
> Subject: Re: [FRIAM] This is scary, and yet very cool...Ai neural networks
> making pictures, look really good
>
>
>
> Oh shit! that's trippy! I though I stumbled over an equally trippy screen
> saver, and (separately) some kind white noise, or background sounds
> generator that used Ai somehow.
>
> Found one of the screen savers that uses Ai. It's free if you let them use
> your GPU and CPU otherwise it's just a few bucks. https://electricsheep.org
> /
>
> I'm not finding the specific whitenoise maker I tried to help with
> insomnia. Just for that side, worked out great. I don't think it was as
> sophisticated as dall-e2. IIRC I had to give it somehelp with some stuff.
> Maybe so it knows where to start? I didn't think about it till now. Because
> if I like backgrounds that are warm and wholesome, it'd need to know what
> mix together and kinds of tones or something? I have no idea. I'm just
> guessing
>
> electricsheep on the other hand can make some stuff that on the entire
> spectrum from trippy, to that's just cheating: a beautiful scenic town kind
> of things.
>
>
>
> Do you know if they needed to train the Neural Networks so it knows whats
> what? like popart from Lichtenstein and andy worhole or what we might find
> at the indian market. Such that later on you say ahah I want a cow print
> slowcooker picture Dall-e can do that?
>
>
>
> On Wed, Apr 6, 2022 at 11:04 AM Marcus Daniels <marcus at snoutfarm.com>
> wrote:
>
> Speaking of which, https://openai.com/dall-e-2/
>
>
>
> From: Friam <friam-bounces at redfish.com> On Behalf Of Gillian Densmore
> Sent: Monday, April 4, 2022 10:01 PM
> To: The Friday Morning Applied Complexity Coffee Group <
> friam at redfish.com>
> Subject: Re: [FRIAM] This is scary, and yet very cool...Ai neural
> networks making pictures, look really good
>
>
>
> 👍
>
>
>
> On Mon, Apr 4, 2022 at 8:37 PM Marcus Daniels <marcus at snoutfarm.com>
> wrote:
>
> I am not a photographer but I have been startled by how recent
> iPhone photos sometimes look better than what I saw when I took it.
> This article explains..
>
>
>
> https://www.newyorker.com/culture/infinite-scroll/
> have-iphone-cameras-become-too-smart
>
>
>
> Filmy-ness sounds like crude interpolation more than structure
> being imposed.
>
>
>
> Suppose I had 1000 pictures of my dog in many lighting conditions
> and from different angles. Using photogrammetry reconstruction
> techniques these could be used to prepare a 3-d textured model of
> her. My iPhone could determine that the 1001st photo also
> included her. It could then reference this model to enhance her
> image in the new context. Maybe inferring the light sources and
> ray tracing her at the required orientation and scale, but at a
> resolution far beyond what was in the photo. That would be more
> art than a photo, but who cares about the truth anymore? Photos
> are to be staged!
>
>
>
>
>
> On Apr 4, 2022, at 5:04 PM, Gillian Densmore <
> gil.densmore at gmail.com> wrote:
>
> Is that also why some of the older software for upscaling
> tricks the new pictures have a kind of saturated or filmy
> thing over them? or is that just from the particular Neural
> Networks or Ai models used? Still very impressive.
>
>
>
> On Mon, Apr 4, 2022 at 5:40 PM Gillian Densmore <
> gil.densmore at gmail.com> wrote:
>
> aaaaaaaaaaaaaaah! coool!! thanks!
>
>
>
> On Mon, Apr 4, 2022 at 5:30 PM Marcus Daniels <
> marcus at snoutfarm.com> wrote:
>
> The idea [1] is that they learn the distribution
> function of different kinds of distortion using a
> machine learning algorithm.
>
> Then that algorithm can invert that distribution
> function. Kind of like a lens can correct for
> nearsightedness.
>
>
>
> [1] https://arxiv.org/pdf/2107.10833.pdf
>
> From: Friam <friam-bounces at redfish.com> On Behalf Of
> Gillian Densmore
> Sent: Monday, April 4, 2022 3:25 PM
> To: The Friday Morning Applied Complexity Coffee Group
> <friam at redfish.com>
> Subject: [FRIAM] This is scary, and yet very cool...Ai
> neural networks making pictures, look really good
>
>
>
> https://github.com/xinntao/ESRGAN
>
>
>
> Stumbled across this looking for a way to gently adjust
> some old pictures of mine without watermarks
> (gigapixel), photoshop wasn't cutting it because not
> enough pixels or data in the originals.
>
>
>
> I am beyond fascinated how do they do it? just guess
> based on colors and add more pixels with that color?
>
>
> .-- .- -. - / .- -.-. - .. --- -. ..--.. / -.-. --- -.
> .--- ..- --. .- - .
> FRIAM Applied Complexity Group listserv
> Zoom Fridays 9:30a-12p Mtn UTC-6 bit.ly/virtualfriam
> un/subscribe http://redfish.com/mailman/listinfo/
> friam_redfish.com
> FRIAM-COMIC http://friam-comic.blogspot.com/
> archives:
> 5/2017 thru present https://redfish.com/pipermail/
> friam_redfish.com/
> 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
>
>
> .-- .- -. - / .- -.-. - .. --- -. ..--.. / -.-. --- -. .--- ..-
> --. .- - .
> FRIAM Applied Complexity Group listserv
> Zoom Fridays 9:30a-12p Mtn UTC-6 bit.ly/virtualfriam
> un/subscribe http://redfish.com/mailman/listinfo/
> friam_redfish.com
> FRIAM-COMIC http://friam-comic.blogspot.com/
> archives:
> 5/2017 thru present https://redfish.com/pipermail/
> friam_redfish.com/
> 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
>
>
> .-- .- -. - / .- -.-. - .. --- -. ..--.. / -.-. --- -. .--- ..- --.
> .- - .
> FRIAM Applied Complexity Group listserv
> Zoom Fridays 9:30a-12p Mtn UTC-6 bit.ly/virtualfriam
> un/subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
> FRIAM-COMIC http://friam-comic.blogspot.com/
> archives:
> 5/2017 thru present https://redfish.com/pipermail/
> friam_redfish.com/
> 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
>
>
> .-- .- -. - / .- -.-. - .. --- -. ..--.. / -.-. --- -. .--- ..- --. .-
> - .
> FRIAM Applied Complexity Group listserv
> Zoom Fridays 9:30a-12p Mtn UTC-6 bit.ly/virtualfriam
> un/subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
> FRIAM-COMIC http://friam-comic.blogspot.com/
> archives:
> 5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/
> 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
>
>
> .-- .- -. - / .- -.-. - .. --- -. ..--.. / -.-. --- -. .--- ..- --. .- - .
> FRIAM Applied Complexity Group listserv
> Zoom Fridays 9:30a-12p Mtn UTC-6 bit.ly/virtualfriam
> un/subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
> FRIAM-COMIC http://friam-comic.blogspot.com/
> archives:
> 5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/
> 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
>
>
> .-- .- -. - / .- -.-. - .. --- -. ..--.. / -.-. --- -. .--- ..- --. .- - .
> FRIAM Applied Complexity Group listserv
> Zoom Fridays 9:30a-12p Mtn UTC-6 bit.ly/virtualfriam
> un/subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
> FRIAM-COMIC http://friam-comic.blogspot.com/
> archives:
> 5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/
> 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
--
----------------------------------------------------------------------------
Dr Russell Standish Phone 0425 253119 (mobile)
Principal, High Performance Coders hpcoder at hpcoders.com.au
http://www.hpcoders.com.au
----------------------------------------------------------------------------
More information about the Friam
mailing list