The Friedman Archives Blog: I'm Swearing Off of Generative AI

This is an impossible portrait to take. A man has adopted most of the neighborhood
cats and feeds them twice a day. They are all very standoffish
and would never gather and pose like this. So I took individual
shots of each cat and created this portrait.

For the past three years I've been experimenting with generative AI - partly to understand its capabilities, partly to stay ahead of the curve (because I don't want to be left behind), and partly because, well, it's amazing. It offers new capabilities never before available to artists. (Very much like how I played with analog synthesizers back in the '70's, or how I played with Steve Jobs' NeXT machine when it first came out.)

I was a "Big Brother" to two fatherless boys for 25 years.
Here's a photo of me and "little brother #2".
I was unhappy with the lighting of the 2nd picture -
it needed some fill and a hair light.

Now I've made the decision to swear off of it. Partly because of the environmental cost (although that will diminish over the next few years), and partly because continuing to share my AI work will surely dilute my future credibility when I present a real image of something that actually happened.

But I do want to share with you how I've been experimenting with it, recent developments, and where this technology will both hurt and help humanity in the near future. Interspersed amongst these discussions will be some examples of how I've been using it. (As always, click on any image to view larger and sharper.)

Much better!

Let's start with the documented negative effects it's had so far:

* The energy cost - This has been making headlines. Electricity costs have gone up in areas close to data centers, which is due mostly to poor oversight and lack of systems-level community planning.

* The environmental cost - The need for massive amounts of water can be misleading. Yes, it needs cooling but you can design closed-cycle cooling systems (like your car's radiator) that recirculates water instead of endlessly drinking it.

* The cost of jobs - Recently there have been lots of publicly-announced job cuts "due to AI". I believe those happened because they wanted to please their investors by demonstrating that they won't get left behind by the coming revolution, and that the company is still primed for growth. But I predict that after 8 months or so many of those people will quietly be re-hired once they realize that AI won't fulfill its promises when it comes to workforce replacement. (Oh, wait... that's already happened.) It will take awhile before the right balance is found; in the meantime we all grapple with the question of where do humans provide the most value?

We attended a wedding in California. Afterward we were
driving the bride's mother to the airport when she
lamented that she didn't get a photo with her and the
married couple. So I grabbed two shots I took and
created one for her. She's now very happy.

* And let's not forget it's replacing photographers. (Also movie makers, set builders, screenwriters, and ad agencies.) And training off of other human creators without compensating them. (Here's where the lawyers will try and save us.)

* Pushback in general - Communities are becoming more vocal about not wanting data centers in their backyards. (Their anger gets justifiably worse when they learn that local governments have approved construction without input from the local communities.) Commencement speakers were getting boo'd so often for promoting AI that it even made the papers.

But AI also presents tremendous benefits in areas having nothing to do with imaging, including the fields of medicine, genetics, material science, and mathematics. Here's a very one-sided youtube video extolling some of these benefits. (And here's a less sensationalist example in the field of immunology.) Here you don't have to worry about hallucinations since every new suggested molecule or vaccine gets tested. This is where the real promise of AI lies.

David Kilpatrick (Cameracraft editor) suggested I do
industrial photography while I was in Uganda so the manufacturer
who provided the equipment could brag about it in their annual
report. Of course there was no time to set it up so I described
how I would have placed the lights, diffusers, and gels. It even
blocked the light from the window.

* New dedicated chips (made specifically to tackle AI workloads) show a lot of promise for reducing power consumption by up to 70% and requiring significantly less cooling. Recently, Nvidia also announced that they can cool chips without water. (Here's another example.)

* Now let's talk about the bubble. Trilions of dollars have been invested in the AI hype, and it's unclear how investors will make that money back. So in the short term AI companies are issuing IPOs to keep them afloat for awhile. (And raising prices.) I've lived through the dot-com crash; and this looks like it might be bigger and worse when the market finally "corrects" itself.

The Cost of Trust

This is the biggest thing for me. It's been a long time since we could believe what we see, but now any idiot can lie for political benefit. You could argue that the Genii's already out of the bottle on that one; but I just don't see any benefit to adding to the problem.

I took the left snapshot with my cell phone. Then I ran it
through AI saying, "Re-create this image as if it was
taken in a photographer's portrait studio".
(I am SO out of business!)

There are also emerging technologies that are trying to address that problem:

* Many Large Language Models are now embedding a Google-invented invisible watermark called SynthID onto all AI output. (It's about time!)

* Others are incorporating Adobe's Content Creation Credentials (which I patented back in the 1990's - just sayin') into both real and generated images as well. This technology is so important it warrants its own future blog post.

* There are chrome extensions that know how to identify the watermark and let you validate an image right in the browser.

It's a start.

* This one really surpised me: Gen Z'ers can spot AI-generated content a mile away and aren't engaging with that content as much. That means marketing teams have incentives to keep hiring traditional art directors (and photographers).

This was a 2024 photo taken in Norway in the worst light possible.
The light blue image was Photoshop's best attempt to clean up the
noise using AI.

This is what happened when I fed the light blue image into AI
and asked it to reconstruct the image as if it were taken with a medium
format camera. Click on this; the detail is incredible.
You can also download the .zip file containing
the above, the original RAW file, plus other people's attempts to
de-noise the originals using various tools.

Compare these 100% crops.

How I've Been Using It

Every wedding photographer's frustration.

I'll be the first to admit that I've been mildly obsessed with this new technology. (Not unusual for me. I played with analog synthesizers in the 1970's because they were new and cool and different.) My friends and family have been inundated with realistic-looking images of people we all know, in situations that never actually happened. I think the appeal for me was that, in the history of the world, it has never been possible to produce such realistic-looking images so easily, and using a computation technique ("diffusion models") that is so unintuitive that the programmer in me simply couldn't fathom doing things in that way.

My experiments get a lot more elaborate. Keep reading!

I made a fake comic book for a family member
whose heros include Dwayne "The Rock" Johnson
and Ryan Gosling.

Then I made some videos.

An early attempt at viideo. Me as #6 from the TV show "The Prisoner"

Me dancing in Sia's Chandelier video

Above is a 30-second clip from my most-watched youtube video,

talking about the 10 best Minolta cameras.

And here's the version where I replaced myself with a gorilla.

Below is the furthest I have pushed this. I wondered if these tools could produce a youtube promotional video for my upcoming book on the Sony A7R VI. So I fed it one of my youtube videos from 4 years ago, and an article about the camera I had made for Cameracraft magazine, and said "Create a promotional video for my upcoming book." It cloned me, it cloned my voice, it wrote the script, and generated all of the insert shots. I had to re-edit it manually to polish the final product, but I'm still ridiculously impressed by what it could do with such minimal input. The important things here are 1) the video is completely synthetic, and 2) in two weeks these tools will get even better.

I have more examples, but I don't want to bore you. :-)

The AI industry is now having it's "Nuclear Power Industry in the 1970s" moment. Large public pushback to a new industry whose environmental costs are climbing and whose benefits and long-term threats are still largely unknown (although the near-term threats certainly seem to be self-evident!)

Other Areas

Gifts for Friends

I think books like mine will become obsolete very shortly - nobody wants to read through an 916-page book when they can just ask ChatGPT the question du jour. I've actually tried training a large language model on one of my books but the book was too large. I'm sure by next year that won't be a problem anymore.

AI Translation is also something I've looked into for my books. Most tools choke on the size of the A7 V book, but I did find a service that seemed to do a good job. Now the problem is someone will have to go through the results by hand and make sure the descriptions use the same nomenclature as Sony's menus for that language.

I was cast in the "Seinfeld" pilot in 1989.
They wrote me out because people in focus
groups kept getting the balding characters
confused.

So why have I stopped playing with this? It is my practice to dabble with new technologies when they become available, then wait a few years to see how the market shakes out and how the technology matures. It will take awhile for the tools to become more power efficient and for society to adapt to this new normal. Swearing off AI now is like swearing off of Photoshop in the 1980's. It may have seemed extreme at the time; meanwhile Photoshop is now considered a normal tool which no longer provokes responses like "OMG THEY MOVED A PYRAMID!". So I'll go back to sitting on the sidelines for awhile and see what develops.

(In my defense, I don't jump onto every new thing. I decided to pass on drones, NFTs and cryptocurrencies, for example. :-) )