Master AI (Artificial Intelligence) Discussion/News Thread

[Sir Sanders Zingmore]

I said no such thing, quite the contrary.

AI is a tool. It's as good or bad as the user that wields it. That's all.

But to claim it's useless is to declare oneself obsolete in many lines of work these days.

To me it is quite unlike any other tool humanity has built.
My fear is that at some point (if not already) the “user” wielding AI will become AI itself (with its potentially completely different notions of “good” or “bad”)

 

[rdenney]

To me it is quite unlike any other tool humanity has built.
My fear is that at some point (if not already) the “user” wielding AI will become AI itself (with its potentially completely different notions of “good” or “bad”)

That has already happened--in many contexts, reality isn't good enough.

Rick "noting Amir's post above" Denney

 

[Sir Sanders Zingmore]

That has already happened--in many contexts, reality isn't good enough.

Rick "noting Amir's post above" Denney

Agree. Hence my “if not already” qualification

 

[xanalog]

perfectly in-focus picture of a random person

What is 'random person' to AI.

 

[pablolie]

To me it is quite unlike any other tool humanity has built.
My fear is that at some point (if not already) the “user” wielding AI will become AI itself (with its potentially completely different notions of “good” or “bad”)

To me it's just another programming language, honestly. It cant do anything we havent programmed it to do, nor can it ever innovate beyond that. Zero chance.

It really can't grow beyond the scope of knowledge it was trained for. Which remains human knowledge and human training of the model for its intended purpose.

 

[Sir Sanders Zingmore]

Zero chance.

I hope you’re right but I fear you aren’t

 

[amirm]

Personally I like the fact I grew up in an age where my pictures reflects the reality of the moment.

I hear you but the camera can be very unforgiving in the way it freezes a moment and at high resolution, allowing the viewer to see things they wouldn't normally notice. This is especially true of very high resolution cameras we have now.

The tool also nicely lets you adjust the level of enhancement.

Such manipulations have been around for decades. Ansel Adams famously said, "the negative is comparable to the composer's score and the print to its performance." He hugely valued post processing in the dark room. We are doing the same here.

Traditionally, it took an expert in Photoshop to be able to perform such optimizations. AI is allowing everyone to be able to do that.

 

[respice finem]

I hear you but the camera can be very unforgiving in the way it freezes a moment and at high resolution, allowing the viewer to see things they wouldn't normally notice. This is especially true of very high resolution cameras we have now.

The tool also nicely lets you adjust the level of enhancement.

Such manipulations have been around for decades. Ansel Adams famously said, "the negative is comparable to the composer's score and the print to its performance." He hugely valued post processing in the dark room. We are doing the same here.

Traditionally, it took an expert in Photoshop to be able to perform such optimizations. AI is allowing everyone to be able to do that.

It's kind of "give and take", OTOH it might be difficult, even impossible to tell true from "AI enhanced" in a few years.
Including, but not limited to, photography.

The more potent a tool is, the more dangerous it can get.

 

[diablo]

Traditionally, it took an expert in Photoshop to be able to perform such optimizations. AI is allowing everyone to be able to do that.

I've had a photography subscription to Adobe for many years, though rarely use it now. Easier to use tools take care of what I usually want to do - and with not using Photoshop much I forget how to use it whenever I do need to use it. So I'll maybe have a look at Evoto.

 

[FrantzM]

Hi

AI are creating a World we are not fully prepared for, nor do we have any clear idea of... but here we go forging ahead .. being our usual hopeful..
...
I am working on a an new initiative. Main issues are Costing and pricing. Before AI became so available and ubiquitous, lengthy discussions with professionals.. It would take days or weeks for a proposal/projectthat would turn a profit for the company in this new field... Trial and error.. mostly costly errors.. Using AI as an assitant.. I arrive in hours at proposal that were taking weeks.. Yes my own experience counts, but I could not have produced these proposals, so quickly without an AI... This elates me and scares me ...

Best.

 

[rdenney]

To me it's just another programming language, honestly. It cant do anything we havent programmed it to do, nor can it ever innovate beyond that. Zero chance.

It really can't grow beyond the scope of knowledge it was trained for. Which remains human knowledge and human training of the model for its intended purpose.

Who is "we"?

That's the question. If I write a program, I'm the "we". If I write a program that can rewrite itself as it measures its own outcomes, then I'm no longer the only "we".

(I'm not defining AI as being merely large-scale data analysis using fuzzy logic, but rather systems that alter themselves as a result of control-loop outcomes that they deem suboptimal with respect to their objective functions. But it can go even further with systems that, like humans, alter their objective functions with changes in context.)

It matters because of legal liability, which I do not think has been worked out at all for an AI world. It's a legal principle that one cannot contract away their liability for negligence. Is it negligent to turn loose a system that can alter itself in ways that create a tort? What if the system no longer allows its creator to exert control? (Don't laugh--this would need to be a requirement for, say, driverless cars, which will lock out external control for all sorts of good reasons, even as it is making a decision unintended by its creator that causes a tort.)

("tort" = wrongful act that leads to civil--as distinct from criminal--liability.)

Rick "might not get worked out until lawsuits start gathering steam" Denney

 

[rdenney]

I hear you but the camera can be very unforgiving in the way it freezes a moment and at high resolution, allowing the viewer to see things they wouldn't normally notice. This is especially true of very high resolution cameras we have now.

The tool also nicely lets you adjust the level of enhancement.

Such manipulations have been around for decades. Ansel Adams famously said, "the negative is comparable to the composer's score and the print to its performance." He hugely valued post processing in the dark room. We are doing the same here.

Traditionally, it took an expert in Photoshop to be able to perform such optimizations. AI is allowing everyone to be able to do that.

Amir, be careful about using Adams in this context. His assumption was maintaining an indexical relationship between the subject and the image, and his "performances" changed tone, boundaries, and occasionally the removal of small details, but it never involved adding components to the image that were not in the original scene (or in some other original scene, even of his own making). It had nothing to do with "truth" in the photojournalistic sense, though that is also a thing. He was always a proponent of being unapologetic about the process of photography, primarily that it renders what is there sharply on the film, as opposed to painting, where the artist had control over what subject elements were included and how they were placed and represented. That implies maintaining an indexical link from subject to image, and was one of the principles of Group f/64 which he founded.

What AI does is more akin to what Jerry Uelsmann did by surrealistically merging images from several sources into a single image. Uelsmann's work has always straddled boundaries, but attempts to identify the boundary it straddles usually lead nowhere. In any case, however, the elements of subject details that he composited together were still his own works, and that seems to me a key point.

(source: Uelsmann.net)

I think it's safe to say that when Photoshop is using material created by someone other than the photographer, and even unknown to the photographer, then the photographer can no longer claim that the work is their own intention. Painters didn't like photography in art museums because they though it was just a tool for people who couldn't draw, but that's an argument of technique, not intentional content. Here, though, when AI puts a "random" face on a photograph--that is, imagery created by someone else with no indexical relationship to subjects the photographer's camera pointed at--it has become a part owner of the art creation process. For utility photography, that may be useful, but the photographer loses some of his rights to claim the work, it seems to me.

If I make a photo of, say, Delicate Arch, which has been photographed 4.895 bazillion times, and then ask AI to spruce it up, the AI engine might well go appropriate a photograph taken from the same location (and all the locations have been used) made on a day when the lighting was better and the clouds prettier, and just replace mine altogether. It would certainly be an improvement. But that would not be a matter of technique, and the indexical relationship between my camera and the scene would be broken.

In this photo (originally made on film), I used Photoshop to remove a couple of tourists that had entered the frame. I could have asked AI to do the same these days. If it did what I did, which is blend the surrounding detail into the images of the tourists, then it maintains the indexical relationship between the subject and the image, even if it is altered. The healing brush in Photoshop does exactly that, and that would fit Adams's concept. But if it went out and found some other photograph to overlay over mine, then that connection is lost.

(Paul Raphaelson used to participate in this forum--he is much more of an expert on this topic than I am.)

Rick "lots of debate in the photography world" Denney

 

[pkane]

That's the question. If I write a program, I'm the "we". If I write a program that can rewrite itself as it measures its own outcomes, then I'm no longer the only "we".

There's more. Most AI and machine learning is based on mapping a huge non-linear function to a data set used for training. The shape of the function is programmed by us, but not any of the details (individual weights or coefficients) as these are tuned using the data set itself. There's no simple way to understand what this function does or even predicting what it will do for any given input after it's been trained.

To say that "we" know what AI will do because we programmed it misses the point that we really didn't program it. The data that was fed into it for training is what programmed it. Most large LLMs today are simply a black box to humans, especially since many terabytes of data were used for training these models.

 

[pablolie]

Hi

AI are creating a World we are not fully prepared for, nor do we have any clear idea of... but here we go forging ahead .. being our usual hopeful..
...
I am working on a an new initiative. Main issues are Costing and pricing. Before AI became so available and ubiquitous, lengthy discussions with professionals.. It would take days or weeks for a proposal/projectthat would turn a profit for the company in this new field... Trial and error.. mostly costly errors.. Using AI as an assitant.. I arrive in hours at proposal that were taking weeks.. Yes my own experience counts, but I could not have produced these proposals, so quickly without an AI... This elates me and scares me ...

Best.

I know what you mean. I always thought of intelligent tools as "automate the trivial, simplify the complex", I used that phrase ad infinitum... but now we can arrive at "automate the complex" too quickly. And many people erroneously use AI tools in a way that just leaves them to do the simple. As in writing technical blogs. Over the last 6 months, I have often fought tooth and nail to make sure blogs have some resemblance of a human tone. AI text is so flat, and while technical articles seldom are leading examples for excellent, artistic prose, we're moving in a direction where both the sheer volume and the quality of text produced turns me off after just a few sentences.

 

[somebodyelse]

There's more. Most AI and machine learning is based on mapping a huge non-linear function to a data set used for training. The shape of the function is programmed by us, but not any of the details (individual weights or coefficients) as these are tuned using the data set itself. There's no simple way to understand what this function does or even predicting what it will do for any given input after it's been trained.

That's where the work on discoverable/explainable AI comes in, and will be required in some cases in the EU so long as the law doesn't get watered down in response to tariff threats. This seems to require external observation of the processing rather than getting the process to 'explain its reasoning' as some recent papers have shown them being deceptive in their explanation - although there's an argument that that's what humans often do too. Older work on image classifiers identified the parts of the image that were critical to the classification, and sometimes turned out to be parts that were in the background, not the object identified.

 

[pablolie]

That's where the work on discoverable/explainable AI comes in, and will be required in some cases in the EU so long as the law doesn't get watered down in response to tariff threats. This seems to require external observation of the processing rather than getting the process to 'explain its reasoning' as some recent papers have shown them being deceptive in their explanation - although there's an argument that that's what humans often do too. Older work on image classifiers identified the parts of the image that were critical to the classification, and sometimes turned out to be parts that were in the background, not the object identified.

The training never stops. I also think it is important to separate between what you could call "do-it-all" LLM models (which often require tens of billions in investment) and smaller, much more wide-spread, specialized inference machines. The latter is very much based on a curated set of data. Where I work, we use it to make complex feature much easier to deploy correctly by customers. Some audio equipment could benefit from that given the vast feature sets...

 

[rdenney]

I know what you mean. I always thought of intelligent tools as "automate the trivial, simplify the complex", I used that phrase ad infinitum... but now we can arrive at "automate the complex" too quickly. And many people erroneously use AI tools in a way that just leaves them to do the simple. As in writing technical blogs. Over the last 6 months, I have often fought tooth and nail to make sure blogs have some resemblance of a human tone. AI text is so flat, and while technical articles seldom are leading examples for excellent, artistic prose, we're moving in a direction where both the sheer volume and the quality of text produced turns me off after just a few sentences.

I would like this ten times if I could.

This is bad enough because only a fraction of extant writing used to train AI is remotely tolerable. Most of it is dull government and academic reports that get published online routinely. Real clarity in the official writings of large organizations remains elusive primarily because that sort of clarity isn't an objective and any form of artistic expression or character is ruthlessly edited out of it. Ask me how I know. So, the bulk of AI's training material has been the worst kind of unexpressive and desultory official writing.

But it's okay, because humans won't be reading it any more anyway. Humans will get a large language model to read if for them and tell them what it says, and never develop the reading skills needed to see deeper meanings. Those summaries will get published, and so the body of writing used to train AI will be fed by an increasing amount of AI writing. One likely outcome is that humans will be less and less able to read what is written with understanding without help from the very tools that facilitated the problem.

Will the writing be better than human writing? On average, probably, at first. But AI has no fear of complexity and when language seeks complex precision to the exclusion of expressive clarity, even with human writers, we end up with legaleze and government-speak. I predict AI-generated documents will surpass legal writing in opacity to any human reader unable to parse out ever-increasing stacks exception clauses.

Many of my colleagues already use LLMs to write their routine reports.

Excellent writers to not spring from the womb--they require training and that usually means writing a lot of desultory crap as part of the learning process. If the AI takes over the duty of writing desultory crap, human writers may never rise above it. Do I fear this? Fear is too strong a word, but one product of being older is realizing that the best possible outcome that is promised is rarely what ends up being delivered. I see proponents predicting a lot of improbable fantasy outcomes, but it seems to me the more likely outcome will be more efficient ways to produce the same mediocrity that we think we are replacing.

Rick "hoping it doesn't turn out to be a zero-sum game" Denney

 

[Timcognito]

If you can't build EVs and tractors anymore, go into AI. The more I read the more I think of the Dot Com bubble that burst in 2000. Back then all you needed was a "killer app" to get investors and maybe that's true for AI? Nice car site if you are thinking of a new car or following the industry

Foxconn Drops Carmaking Bid, Sells Lordstown Factory for a Total $375 Million - TFLcar

After failed bids to manufacture EVs, Foxconn sold its Lordstown, Ohio plant to an "existing partner" for $375 million.

tflcar.com

 

[pablolie]

... The more I read the more I think of the Dot Com bubble that burst in 2000...

That's most definitely true. The billions and billions that go into startups with unexperienced leadership is staggering, and losses are sure to happen when this moment of over-inflated expectations fades and reality sets in. I get invited to sales pitches to our company from several of these startups and the vast majority are utterly and ridiculously bad.

 

[sam_adams]

There's no simple way to understand what this function does or even predicting what it will do for any given input after it's been trained.

And therein lies the problem with the current GPT systems. If the programmers who wrote the code don't understand why it does what it does; how can we as users/consumers trust it to the point where we are certain that we won't become victims of the "mistakes" that it will inevitably make? We can be certain though, that if the programmers can't figure out what's happening now with their code, when the systems become self-modifying, the likelihood that they will loose control of the system increases dramatically.