That’s not enhancement, it’s removing our reference for when people actually looked like people. I just posted in the unpopular opinions community a week ago complaining how nobody in movies looks human any more. I’m going to be really disappointed if they go back through all the old movies with an Instagram filter. Come on, motherfuckers. It’s okay to look human.
Pretty sure these tools are often seeded with prompts that enforce diversity. Bing does the same or similar. I’m more amused by this, as the process isn’t aware and can’t actively enable or disable these settings.
To actively fit a historical prompt, it would need to not only consider images from the period, but also properly synthesize historical data to go with the prompt.
That would require some kind of machine capable of learning, a model of language so incredibly large that it can comprehend these linguistic nuances, or an intelligent form of artificial device.
Wonder if we’ll ever have something like that in the future.
There’s a Sci-fi horror book I enjoyed, called “John Dies At The End”, that posits an alternative history in which computers were created from the brains of pigs.
As a consequence, the civilization is heavily invested in harvesting organs in the same way that we’re invested in drilling for oil.
Yes, I saw some talk and a screenshot somewhere that showed that apparently in its current state, Gemini can (or could) be asked to output the prompt enhancements it used along with the generated images.
The screenshot showed someone asking for images of fruit, and the enhanced prompt included “racially diverse groups of people”. Now if they’re inserting something like that even for images containing no people at, it stands to reason that this is just a default enhancement they ALWAYS apply, no matter the prompt, which would explain the racially diverse Nazis (and all the other brouhahahas we’ve seen from them).
I…what? Who thought this was a good idea, or even a thing that was needed? Why would you need to ‘modernize’ historical artwork? Why the great flying fuck would you do this by putting it through an AI program that - extremely fucking crucially - changes all of the minor details of the piece? This is absolutely terrible. Whoever worked on this needs to unplug all of their devices and go peel some potatoes.
Check out my milk bread posts if you’re into potatoes.
Can you share more about what information you think is lost in the high fidelity generations like the lighthouse and colossus, especially given you can toggle the images?
The fidelity with the lines and subject is pretty high right now via gpt4 vision and controlnet, but I could work on getting it higher – Which images bother you the most and what level of fidelity/auto restoration would cause you to have a more positive reaction?
My problem is the basic concept. These are historical documents, and you are tampering with them. It’s like translating the Declaration of Independence into leetspeak. As a one-time gag, it’s worth a chuckle, but the idea that it would be an ‘improvement’ or a ‘modernisation’ is an insult. The ‘fidelity’ of your process is irrelevant. You do not and should not need an artifical ‘higher resolution image’ of a centuries-old painting.
That’s what I thought – It seems like there’s a group of people this bothers and aren’t interested in regardless of the outputs. It feels ideological, but I won’t go that far and claim that for you or them – Anyway, this isn’t made for you and it is OK that we feel differently.
Thanks for sharing your feedback and opinion again! Have a great day!
What are you doing? That painting shown on Wikipedia for the lighthouse of Alexandria is an actual painting, you can’t just replace that with another random image and say look this is better. The painting is the painting, with its flaws and character and historical aspects.
Yeah I meant the drawing, the original. I would consider that a painting, but that’s semantics.
One of the biggest issues I see with this is the two people in the bottom right. They are naked and most likely slaves. One of them has a lighter skin color and the other a darker skin color. Whilst it isn’t the focus of the original drawing, it is a part of it, because it is a part of history.
In the new “AI improved” version, not only is all detail lost and are there straight up mistakes, the light colored slave now has clothes on (implying it’s not a slave) and the darker colored one is simply deleted. Rewriting history by deleting slavery using AI is a HUGE issue and highlights why you should absolutely not do this.
The original contains so much more information and context, you are deleting all of that. Basically everything that made the original worth anything and worth including in the article is gone.
If you say you can’t see the differences you are full of shit, get the fuck out of here.
No need to be ugly I don’t think — Thanks for the feedback all the same.
I honestly did not notice it because my focus was on the lighthouse structure itself and that level of fidelity wasn’t something I personally needed as an addendum to the original image (not replacement)
So, I could get tiny subject fidelity up by using larger images from the start, but that would dramatically increase generation time.
Right now an image runs through GPT AND generates in about 10-15 seconds. If we pretend you were using this, what the max time you’d want to wait for the highest quality image?
If I added an upscale button to the AI image that produced a 4X much higher detail/fidelity image but took longer, would that be a solution?
I think you have a flawed understanding of what “AI” is and does.
It doesn’t enhance, it doesn’t improve, it doesn’t increase tiny subject fidelity. It makes stuff up, it invents data, in other words it’s total BS.
Using something like AI upscaling for videogames is fine, because if it fucks up it’s at worse a tiny glitch to get annoyed with. Using it on something like Wikipedia, which for many people is a source of information, is VERY dangerous and downright stupid. You can’t rely on anything produced by AI. It isn’t the magic zoom and enhance button we know from TV and movies.
When Elden Ring came out, I wanted a huge ass poster of it. But even the official press release only included images of limited resolution, fine for a wallpaper on the computer, not fine for a high quality print. I messed around with different AI upscaling techniques till I found one I was happy with. Even then I spent hours tweaking the parameters and throwing a lot of computing power against it, till I got out something I was happy with. And even now I know small little details which aren’t right because of that algorithm, but I’m the only one who knows or sees so I was OK with it.
If you are learning about subjects by using AI, please stop and use actual primary sources. What you are learning is fiction, a fantasy and not real life.
So, this is more than just running the image through an AI generator – It uses a few techniques to lock the image to the lines in the original image and the prompt is aware of the article data and the image via GPT vision, controlnet, and the internet.
Also, this does not replace the images on Wikipedia, it’s a chrome extension that let’s you toggle between original and generated images – The intent is if you see an old 1700s etching and wonder what it really looked like – Or see a poorly drawn Mughal era painting and wonder what the scene might have looked like in real life – The only real ‘funcitonal’ use I’ve seen building it is with coins and other things that are ‘worn down’ It does a pretty good job at making that stuff more visible – There’s a few coin examples in the post.
Can you look at the line drawing of the lighthouse of Alexandria and the AI generated image for me and tell me if there’s some level of fidelity improvement that could be present to make you feel differently? I struggle to find a lot of differences other than the color.
The ‘upscale’ button could just let us start with a higher resolution starting image with all details preserved – In the painting of the lighthouse, where the boy is removed, that kind of thing would get fixed and small characters would be much better preserved, at the cost of generation time – I’m not saying just upscale the AI image.
On the comment about fiction/fantasy – The majority of the images we’re modifying are not ‘primary sources’ in that Hermann Thiersch never saw the Lighthouse – This feels like the same level of fantasy since we’re using his original image with such high fidelity. I’m curious to get your thoughts.
Please just stop, you don’t know what you are doing.
People who went to school for over a decade in this subject would be able to tell you a thousand things about some of the images you are referencing. People worked hard to include the best possible image with the article.
You then go and generate some BS image and say: “I struggle to find a lot of differences other than the color.”
And no these things cannot be fixed, there is no fixing a flawed principle. You can’t fix it by renaming it or by saying it’s only a chrome extension. Please stop.
Do you have any articles or reading I can do on what those ‘thousand’ things would be? I can definitely build that into the model either with fine-tuning or connecting GPT to the internet.
I wholesale disagree things can’t be fixed and your logic there doesn’t really track. In general your manner reminds me of the famous Sartre quote. You don’t seem to really be interested in engaging in good faith. I find your failure to even attempt an answer at my question suggests your true motives.
If you press them too closely, they will abruptly fall silent, loftily indicating by some phrase that the time for argument is past.”
Oh, I see – I meant the 2nd lighthouse picture. The one that is much higher fidelity to the original image. I see what you’re saying about the first image
The special was actually really good in my opinion. I personally like the idea of preserving cultural icons a la the talking heads in Futurama via AI. I do think the estate should get royalty rights like they would with deceased artists, but why not embrace this medium of immortality?
the first two minutes or so of the special make it abundantly clear it’s an imitation, it’s intended as an impersonation, and should not under any circumstances be taken as a “AI clone” of carlin. they’re very up front about it, and the reasoning for doing it.
it’s been blown way out of proportion by newsfeeds jumping on the “his daughter doesn’t like it” side of the story. it’s no different than comedians doing impersonations of each other on late night talk shows
My major problem with it is not getting permission from his heirs before creating it. With that being said, the AI Bill Cosby joke was pretty damn good.
isn’t it weird that it’s not them, it’s like a caricature of them? And, as time goes on is their legacy who they actually were or would it be this? It seems questionable even for people who agree before they die, but it seems really unethical for anyone who didn’t agree or were even aware of the possibility.
Preservation would be keeping copies of his standup. This is not a special by George Carlin, this is a copycat. And the copycat will get stale because it won’t create anything new, it will only regurgitate jokes based on the very limited material that exists.
I am using ChatGPT4+ with the code interpreter and I am finding it closer to 90% accurate writing 50-200 lines of c# code in Unity. Beyond 200 it starts to have more issues and the accuracy drops. It has saved me so much time refactoring my project.
Yeah. They buried it in there (and for some of their experiments just said “ChatGPT” which could mean either), but they used 3.5 and oddly enough, 3.5 gets 48% on HumanEval.
They “burried” it in the methodology section, where they describe how they generate prompts. This is the place I expect this to be mentioned, or am I missing something? Where else would they put it.
It’s a pretty important fact since there’s a huge difference between 3.5 and 4. Mentioning it once in one place is not great, plus they also just mention ChatGPT without specifying 3.5 or 4 earlier in that paragraph. The problem I have is this has led to press (and hence many other people) thinking ChatGPT is terrible at coding when in fact using the GPT 4 version, it’s actually pretty decent.
Recent work claims that large language models display emergent abilities, abilities not present in smaller-scale models that are present in larger-scale models. What makes emergent abilities intriguing is two-fold: their sharpness, transitioning seemingly instantaneously from not present to present, and their unpredictability, appearing at seemingly unforeseeable model scales. Here, we present an alternative explanation for emergent abilities: that for a particular task and model family, when analyzing fixed model outputs, one can choose a metric which leads to the inference of an emergent ability or another metric which does not. Thus, our alternative suggests that existing claims of emergent abilities are creations of the researcher's analyses, not fundamental changes in model behavior on specific tasks with scale. We present our explanation in a simple mathematical model, then test it in three complementary ways: we (1) make, test and confirm three predictions on the effect of metric choice using the InstructGPT/GPT-3 family on tasks with claimed emergent abilities, (2) make, test and confirm two predictions about metric choices in a meta-analysis of emergent abilities on BIG-Bench; and (3) show how similar metric decisions suggest apparent emergent abilities on vision tasks in diverse deep network architectures (convolutional, autoencoder, transformers). In all three analyses, we find strong supporting evidence that emergent abilities may not be a fundamental property of scaling AI models.
Page two of the paper states their thesis pretty succinctly:
In this paper, we call into question the claim that LLMs possess emergent abilities, by which we specifically mean sharp and unpredictable changes in model outputs as a function of model scale on specific tasks. Our doubt is based on the observation that emergent abilities seem to appear only under metrics that nonlinearly or discontinuously scale any model’s per-token error rate.
And figure two walks through the argument for it concisely as well:
(OCR for image: Figure 2: Emergent abilities of large language models are creations of the researcher’s analyses, not fundamental changes in model outputs with scale. (A) Suppose the per-token cross-entropy loss decreases monotonically with model scale, e.g., LCE scales as a power law. (B) The per-token probability of selecting the correct token asymptotes towards 1 with increasing model scale. (C) If the researcher scores models’ outputs using a nonlinear metric such as Accuracy (which requires a sequence of tokens to all be correct), the researcher’s measurement choice nonlinearly scales performance, causing performance to change sharply and unpredictably in a manner that qualitatively matches published emergent abilities (inset). (D) If the researcher instead scores models’ outputs using a discontinuous metric such as (Multiple Choice Grade, which is similar to a step function), the researcher’s measurement choice discontinuously scales performance, causing performance to change sharply and unpredictably in a manner that qualitatively matches published emergent abilities (inset). (E) Changing from a nonlinear metric to a linear metric (such as Token Edit Distance), model shows smooth, continuous and predictable improvements, ablating the emergent ability. (F) Changing from a discontinuous metric to a continuous metric (e.g. Brier Score) again reveals smooth, continuous and predictable improvements in task performance, ablating the emergent ability. Consequently, emergent abilities may be creations of the researcher’s analyses, not fundamental changes in model family behavior on specific tasks.)
But alas, AI/ML isn't my wheelhouse, so the paper quickly starts to go over my head.
You don’t own the rights to the voice of every actor who arguably sounds kinda like you. OpenAI had an idea of a type of voice they wanted, when Scarlett said no they hired a voice actor. I mean, what? There are many valid criticisms of Sam and OpenAI don’t get me wrong but this is one I just can’t get on board with.
I feel like I’m missing something though because so many people are commenting on this as though Scarlett had her own voice used without her consent or something.
This made me think a little too, but then I started thinking about how people talk. Even if a person’s tone is similar, the mannerisms are still drastically different. The voice actor had to spoof Scarletts voice well enough to even fool Scarletts friends.
While I haven’t read Altman’s tweets (or tweets from someone else at OpenAI?) personally, rumor has it he knew what he was doing with a specific voice actor. The intent was to spoof her voice. The intent is probably more damning than the actual act, TBH.
To summarize, there are a few nuggets here: They approached SJh first with a specific reason to use her voice; Some asshat bragged about spoofing the voice on Twitter; There was clear intent of generating a likeness of SJh.
(Thinking about the broke actors for a second… /s) Their face and voice are at the core of their career, similar to how company branding is makes a company unique. While I am not a lawyer, it seems there are some parallels with trademark and copyright law here.
“Accidentally” using a voice that sounds like SJh would be a really poor argument now as well.
This made me think a little too, but then I started thinking about how people talk. Even if a person's tone is similar, the mannerisms are still drastically different. The voice actor had to spoof Scarletts voice well enough to even fool Scarletts friends.
I don't get this. Why are you assuming they constructed the voice with only the samples from another voice actress and didn't use any from Johansson? Why are you assuming they used the samples from that voice actress at all and didn't only use samples of Johansson's voice they scraped from all corners of her prolific history of work?
Any random company I would give the benefit of the doubt, but these AI companies have specifically shown they don't care about copyright law specifically or ethics in general, and they definitely have no qualms lying about where they get their data and what they do with it.
If parts were generated, copied or even duplicated it seems somewhat irrelevant. They intended to, and did, generate SJh’s likeness, by whatever means, and that is the key point.
But yeah, they probably mixed and matched voice samples to their liking. I wouldn’t doubt that for a second. If actual samples were used in the final product, that would be extremely damning.
They reached out to her a couple days before they launched, and said hey do you want to maybe reconsider that thing where we asked you about this a couple times, and both times you told us to go fuck ourselves
And then they told the media that they were in discussions with her, when the discussions were her lawyer telling them to go fuck themselves
And then Altman tweeted "her"
And then when it launched, it was according to her so freakishly similar that her friends were weirded out by it
If it was some different actress saying hey this sounds a lot like me, that wasn't the one that they clearly had in mind when they were making their plans about it, then I could see a pretty strong argument to say hey relax buddy sometimes different people just sound similar
I don't really know and I don't care enough to listen to samples for myself and see what I think. But just based on the above set of random facts I feel like probably she has a fairly strong case.
This the same guy who (allegedly) has apparently been having unethical CNC kink parties💀? Like no judgment, I’m actually kind of into CNC, but you have to Actually get enthusiastic informed consent first (and during) and use safewords. You can’t just foist it on people who happen to show up to a party, that’s literally (allegedly) just sexual assault. Seems like he might just be (allegedly) completely indifferent to the concept of consent.
Apparently the person didn’t specify any specific OpenAI executives by name, so it’s uncertain whether Altman knew about the parties. Still reflects poorly on him if this is the culture at OpenAI tho imo.
artificial_intel
Hot
This magazine is from a federated server and may be incomplete. Browse more on the original instance.