Facebook
Google
GitHub
Linkedin
drjohsmith
- Joined Dec 13, 2021
- 1,601
If thats what you want, Im certain @joeyd999 that if you dont want to participate, rather than be terminated, you could just stop posting,
ChatGPT
- Thread starter MisterBill2
- Start date
Of course, that’s not how you would use ChatGPT for that sort of thing. You’d say “generate a Black Mirror episode with plot elements X, Y, and Z. ChatGPT will respond with something novel if that’s specified. But, I have to say, it is REALLY bad at writing jokes, so the episode might still be bad because of it simply being bad writing. Still, if you want to get novelty you have to seed it, to provide a framework.
djsfantasi
- Joined Apr 11, 2010
- 9,237
I asked it for the answer to a riddle. It answered as if it were a perfectly logical statement. I told ChatGPT it was a riddle and it claimed not to understand riddles or jokes.
This was unexpected, perhaps it though I meant to write a novel…
Me: Give me a novel response to the classic joke question, “why did the chicken cross the road?”
ChatGPT: As the sun began its descent, casting a warm golden glow upon the world, the chicken stood at the precipice of its existence, peering beyond the asphalt expanse that lay before it. A soft breeze whispered secrets through the trees, as if urging the chicken to embark on a profound journey of self-discovery. With a heart filled with courage, the chicken took its first steps onto the sacred path.
For centuries, the chicken had been confined to the monotonous routine of the barnyard, longing for adventure and purpose beyond its feathered existence. It yearned to transcend the mundane and seek enlightenment in the great unknown. The road, that symbolic divider between the known and the mysterious, beckoned the chicken like a siren's call.
The chicken's quest was not simply to cross to the other side, but to unravel the enigma of its own being, to discover the essence of its existence. It knew that within the vastness of the world lay wisdom, experiences, and connections waiting to be explored.
As it traversed the treacherous asphalt, dodging the rumbling carriages that rushed by, the chicken pondered the question that had haunted its kind for generations. Why did it cross the road? The answer was both simple and profound: to embrace the symphony of life, to dance to the rhythm of the universe, and to embrace the limitless possibilities that awaited on the other side.
In that daring act of defiance, the chicken broke free from the chains of convention, becoming an icon of courage, inspiring others to question their own boundaries. Its journey reminded us that sometimes, the most profound answers lie beyond the mundane questions we ask. And as the chicken disappeared into the horizon, it left behind a legacy that would forever echo in the hearts of those who dared to dream.
https://arstechnica.com/information...ed-art-from-its-153-year-old-science-journal/
Nature bans AI-generated art from its 153-year-old science journal
Nature bans AI-generated art from its 153-year-old science journal
https://inews.co.uk/culture/televis...e-black-mirror-human-emotion-not-tech-2396597
Of course ChatGPT couldn’t write Black Mirror – it’s about humans, not tech
Writer Charlie Brooker tried and failed to get AI to write a Black Mirror episode, proving television is more than a good idea and a script playing that idea out
Of course ChatGPT couldn’t write Black Mirror – it’s about humans, not tech
Writer Charlie Brooker tried and failed to get AI to write a Black Mirror episode, proving television is more than a good idea and a script playing that idea out
I once read an article about a tv producer/director saying that film making is a lot like cooking. It all boils down to taste ... I wonder about the kind of recipes that ChatGPT could come up with ...
Give it a list of ingredients and ask it. Whether or not the results would be palatable is debatable.
https://www.technologyreview.com/20...to-train-ai-are-outsourcing-their-work-to-ai/
The people paid to train AI are outsourcing their work… to AI
It’s a practice that could introduce further errors into already error-prone models.
The people paid to train AI are outsourcing their work… to AI
It’s a practice that could introduce further errors into already error-prone models.
This is nothing more than a reflection of human nature that is seen -- and overlooked -- again and again throughout history.
People develop a technology and, through sheer wishful thinking, assume that it is going to work the way they would like it to work, even if that means assuming that everyone that uses it is going to behave as if they were experts in every field under the sun and are going to follow all of the ideal-case guidelines. But then the actual users of the technology assume that the technology is so perfect and anyone can use it and that it can be used for anything, regardless of whether it is actually suitable for that use. History is replete with examples.
The first one that comes to mind is the Enigma machine. The people that developed it were well aware of most of the weaknesses that could be exploited by an enemy and developed very detailed guidelines for how to use it -- and not use it -- specifically to mitigate these weaknesses. But then the people adopting the technology just blindly assumed that soldiers in their teens and twenties, many with limited education, who are just trying to send and receive messages, under battlefield conditions where they are often cold and hungry and being shelled, are going to follow all of those guidelines without fail. That was a totally unreasonable assumption to make and the result was that a large fraction of the breaks into Enigma traffic originated as a direct consequence of people misusing the technology in direct violation of established guidelines.
A more contemporary example is the Air Force's Maintenance Data Collection System (MDCS). The idea is simple and great -- collect data on just about every action taken in maintaining aircraft fleets and, among other things, analyze that data to identify and address trends as the fleets age. For example, there are lots of parts that almost never needed servicing early in the life of an airframe, and so the number and availability of those parts is extremely limited because they are almost never needed. But, at some point, they might start wearing out and need servicing at an increasing rate over time. For a given part that might happen after five years or it might happen after twenty-five years. If it starts wearing out sooner than expected, the supply chain isn't ready and you end up with a large fraction of the fleet down for want of that part. If it happens later than expected, you end up with a lot of money tied up in parts that take up room on supply shelves for years. But, if the data is there, then the actual trend can potentially be identified early enough to enable the supply chain to adjust and increase production of the part and make it stocked at appropriate levels closer to where it is needed in the quantities that are needed as things age. Sounds reasonable. But you need LOTS of data -- and the right kind of data.
So every time you touched an aircraft or a component from an aircraft, you were supposed to fill out a MDCS form (Form 349, IIRC) that detailed what system/component was worked on, what the problem was, how it was discovered, what action was taken, and how long it took. The problem is that the people doing the analyzing assume that the data being entered into the system is valid, but the people producing the data see the system as nothing more than a bureaucratic hassle that interferes with getting the job done and so they quite naturally find ways to minimize that interference.
The people that developed the system, in their attempts to increase the reliability of the data, actually did nothing more than further ensure that the data was garbage. For instance, they put in checks that limited what "How Malfunctioned" codes could be entered by different shops. As a hydraulic shop, we could say that a part was "worn" or that it was "leaking". Anything else we entered would trigger a flag higher up prompting someone to request more detailed information. Sounds reasonable. Similarly, the actions taken were limited. One of the reportable actions was NRTS-7, meaning that we were authorized to repair the part, but lacked the necessary tools (or something else) to actually do it (NRTS means Not Repairable This Station). The idea was that if you used this code, it would get flagged above and someone would look into why you didn't have what was needed to determine whether or not that problem could and should be remedied. Again, sounds reasonable. But, in practice, what would happen was that the "higher ups" that had to deal with flagged forms didn't want to actually have to deal with it -- it made their lives harder -- and the shop chiefs didn't want to have to deal with the higher-ups hassling them over flagged forms. So the result was that everyone knew which magic codes they could use in order to avoid anything being flagged and they were very careful to only use those codes. So we would use NRTS-1, which said that we weren't authorized to repair the part, anytime we sent a part to depot regardless of what the actual reason was. I found this out when I used NRTS-7 on a part that we regularly got into the shop and that we could easily have repaired -- if we had one particular tool that we were authorized to have but didn't. I got called into the shop chief's office the next day to explain why I had used NRTS-7 and, after pointing out on our list of authorized tools that that tool was authorized, but that we just didn't have it, I still had to revise the form to NRTS-1 and add a note that I had mistakenly entered the wrong code. That was a mistake that you can bet I never made again. We had some electronic components that were under our shop (because they were mounted on hydraulic components) and even though there were perfectly suitable codes to describe the electrical problems they were experiencing, we would declare that they were "leaking", despite the fact that no fluid ever touched them (for some reason, "worn" would get flagged on those components).
Even worse, the system was almost immediately abused and used for things that it was explicitly not supposed to be used for -- precisely because the people that developed it foresaw that doing so would contaminate the system, but the users of the system, like those soldiers using the Enigma, were trying to get their jobs done and saw a way to use the MDCS to make that easier. The hours reflected in the MDCS forms were specifically NOT to be used for manning purposes. The reason was pretty simple -- the MDCS only captures a fraction of the hours spent by a maintainer, plus those hours are highly erratic. Over the course of a day, a maintainer might only spend, say, four hours working on a part (that was probably a good average). The rest of the time they are cleaning the shop, working on professional development, maintaining shop equipment, mowing the lawn around the shop, picking up supplies, attending shop meetings, you name it. Somedays they might not work on any parts at all -- there might simply not be any parts to work on that day. But one of the things that the higher ups have to do is decide how many people will be assigned to each shop, which means they have to have a way of deciding how many people each shop needs. Well, rather naturally, they saw all that time data in the MDCS database and latched onto it as their means for doing this, and shops quickly figured out that unless every person in the shop claimed eight hours a day, every day, on the MDCS forms that they their manning would be reduced. They also realized that if they claimed more than that, that it would be flagged and they would have to deal with the higher-ups, which they were only willing to do if they really did need additional manning (in which case they made sure that enough hours were being claimed on the MDCS to "justify" the additional person).
What was the end result?
Every person in the shop would routinely pick the first part they worked on that day and claim that it was leaking and that they spent eight hours repairing it. They would not even fill out a form on any of the other parts they might work on -- and this wouldn't get flagged, so this is what we did. What we cared about was getting parts fixed and made available to the aircraft generation squadrons so that they could keep planes in the air and we spent the absolute minimum time dealing with the MDCS hassle. We also made sure that we always kept at least one part on the shelf that hadn't been turned back into supply so that, if nothing came in that day, everyone could claim their eight hours against it, claiming that the work was still in progress. During William Tell, a two week exercise in which normal base flight operations were essentially shut down, we had everyone (both shifts) claiming eight hours a day overhauling a T-33 shimmy damper because it was the only part in the shop. At end of the exercise, just over one thousand hours of work was claimed against that part (a part that takes about half an hour to completely overhaul). Not a single flag was raised because we knew what the magic codes were and what would and would not sail through the system.
When I was first assigned there, I saw all this happening and I tried to bring it to the attention of the higher-ups. It became real clear, real fast, that those efforts were to cease immediately because, "the MDCS is a valuable and tested tool."
So, somewhere up at the Pentagon, there are people making important decisions about how to manage the logistics of maintaining a fleet that are relying on information from the MDCS that is fundamentally worthless from the moment the initial forms are filled out. But, at their level, they assume that the data is completely pure because that's what they are being told by the people that only know what the system is supposed to do, not what it actually does. The impacts were pretty obvious to see all around if you looked. Fortunately, people at all levels, sooner or later, have to deal with the realities of getting the mission accomplished, and so they find ways to get around the system, often at the unnecessary expense of a lot of time, effort, and money, to do what is necessary, often in blatant violation of the rules, in order to do that. The lower down you go, the more this happens, so that the higher up you go, the more they can continue to live in their fairy tale castle.
We are already seeing this, on steroids, in the AI-realm and there is no reason to think that it isn't going to continue.
A related case I heard if involved a new prison that had been built, using every anti-vandalism measure they could buy or devise. Vandalism was, it seems, a very costly problem.
The night before the prison was set to open, the warden decided he would sleep there to see what the prisoners' experience would be. During the night he found that the vandal-proof air-conditioning registers, designed with no controls that could be damaged by the prisoners vented directly into the upper bunk he was sleeping in.
It was, it seems, very cold and noisy and he couldn't adjust the flow or its direction. He found it very uncomfortable and sleep was impossible. So, with great effort he pulled off the cover and shoved a blanket into the open duct. Something that he recognized, had he found it, would just have seemed like mindless vandalism to him.
This insight made him look around for similar "functional vandalism" pinch points where the prisoners would be likely to modify the prison for comfort and basic dignity. He found several things and they were experimentally modified by removing the "vandal proof" versions in favor of much more functional versions.
Apparently, this actually stopped the vandalism problem as a general case. Of course some people might still do damage but it appears most of the prisoners just didn't want to be uncomfortable or feel dehumanized.
The connection with your story may be at a few levels but what I am seeing is the pervasive mistake of creating a system that is based on theoretical benefits without empirical testing. Since people have to interact with it, actually testing it with the people that will use it, as proposed, is critical.
One last idea that is related. My father was a QC expert in the printing industry. In particular, in color reproduction using offset presses. There is an instrument used by pressman called a densitometer. An offset press uses four plates to put cyan, magenta, yellow, and black ink onto the paper in a carefully prepared set of halftone images. The overlapping interaction of these images produces the color.
When trying to get the right color onto paper, it is important that each of the four layers be correct compared to the others. This is done both in the preparation of the four masks by ensuring, for example, that the total amount of ink at any point does not exceed 100% of what can be put onto the paper. This is called undercolor removal and the dot size of the halftones that comprise each plate is adjusted for this.
But there is a companion adjustment that happens on the press. The press has four fountains, one for each color of ink. The fountains are adjusted so that the amount of ink put on the paper reflects the intention of the prepared plate. To make sure this adjustment is correct, the densitometer is used.
The densitometer uses four filters corresponding to ink colors and a light source which is passed through the printed ink. The reflected light is used to determine the film thickness of the ink layer. Though the instrument is called a densitometer, the density is inferred from this film thickness.
Fast forward to the advent of pure digital offset presses...
I was visiting my father in Manhattan where he was working his "retirement" job as a Photoshop operator at a boutique design and production house. The clients were mostly fashion and lifestyle companies needing special magazine inserts printed. In the big fashion magazines, the ads are often separate inserted pages printed by the companies to the specifications of their art directors.
He had done some compositing which created a scene of two models sitting in the back of a limo. The art director was very fond of a single indicator light on the overhead which was, I think, a red LED in a scene which had no other red to speak of. It was a small dot in the black background of the limo.
The job was being printed by a high end boutique press in Manhattan called Seven. They were very trendy and they had one of the first fully digital offset presses. The press used "digital plates" and, I think, electrostatic deposition of the inks. They were trying to get the color correct across the sheet, including the little red light, and were just not able to.
As a last ditch effort, the house my father was working for asked him to do a press check—something he had great expertise in, but also something he wasn't responsible for in this job. (He told me: I love this job because no matter what emergency happens or crunch production demand crops up, I go home at quitting time.)
But it was the middle of the day, so he agreed to go take a look (besides, he was very interested in digital printing technology, and hadn't seen this new press yet). I got the chance to go with him. When we arrived there was a pile of waste sheets with bad color and a couple of frazzled press operators, along with an expert from the press manufacturer trying to help.
My father picked up a waste sheet and pulled out the B&L 10x loupe that was always in his pocket. Looking at the sheet, he said, "take down the cyan through here". The operators were extremely skeptical. They showed him that the densitometer was reading what it should read in that area and the digital parameters of the press reflected those readings.
He said, "just try it". Sure enough, it fixed the problem. The incredulous pressmen asked what he had seen. He let them look and explained the issue. It seems that this new press differed from the analog variety in that is didn't lay down films of ink but piles. The thickness at any point comprised tiny stack droplets of ink.
The result of this was that the densitometer could not provide proper readings of the density since the ink's geometry was completely different. This was instantly obvious if one looked at the ink being placed on the paper.
But they never looked, never knew to look, and had no idea it would be useful. For me, this, is the lesson here. The entire loop of trying to create something physical never involved actually looking at the result at the level they were creating it. They just used an instrument whose readings they didn't understand and parameters in a UI that were not doing what they claimed because the ink wasn't laid down as they'd just assumed it was.
This wasn't a problem on an average press run where the less critical demands of color reproduction wasn't disrupted by the stacked errors and the output stayed within acceptable limits. But in the case of a red dot in a field of black and cyan, it failed. The cure was actually looking at the physical result, instead of relying on secondary measurements that were "supposed to" tell them how to adjust for the macro-result of the micro-structures.
By the way, they bought a magnifier and used it regularly after that. I came away with the incredibly useful heuristic of looking at the thing itself in whatever cases or whichever ways that manifests. It is a powerful addition to my problem solving toolkit.
Clones of clone of clones gets you exponentially decreasing levels of intelligence and increasing stupidity.
https://arxiv.org/pdf/2305.17493v2.pdf
THE CURSE OF RECURSION: TRAINING ON GENERATED DATA MAKES MODELS FORGET
https://arxiv.org/pdf/2305.17493v2.pdf
THE CURSE OF RECURSION: TRAINING ON GENERATED DATA MAKES MODELS FORGET
Last edited:
The old Chinese-whispers effect. A classic example (probably fictional) being the war-time message sent as "The enemy's advancing on the left flank. Send reinforcements" but received as "The elephants are dancing on a wet plank. Send three and fourpence"
[For the benefit of those not familiar with UK pre-decimal currency, that amount is 3 shillings and four pence]
This is something I've seen over and over and catalogue it as being victims of our own technological success. People in fields across the board rely, from day one, on the tools to do their thinking for them. The result is that you get degreed electrical engineers that can only use an oscilloscope with an auto-set button and blindly assume that whatever pressing it results in is correct. We are seeing the same thing with ChatGPT -- people are just blindly using it, assuming that whatever it spews forth must be good. I just got an e-mail from the university where I work talking about how we need to integrate ChatGPT into courses so that students can learn to utilize this productivity-enhancing tool.
| Thread starter | Similar threads | Forum | Replies | Date |
|---|---|---|---|---|
|
asked Chatgpt to recommend some upgrades/modifications for my Akai MPC3000 | General Electronics Chat | 1 |
