Why has humanity suddenly decide to spend enormous amounts of resources on tech that tries to create stories? Stories that MAY, occasionally relate to reality, but mostly just to create fantasy. Saturday morning cartoons used to amuse me when I was 6 years old, and I guess a lot of adults are amused by moving cartoons about fictional characters. We are also amused by motion pictures of real adult actors, by songs and music performed by people, and by printed words that tell fictional stories (novels). There seem to be plenty of people available to generate this amusement, so why do we need to created tech that creates it?
Because there's still plenty scarcity when you want specific artwork or songs, and bringing the cost of that down a few order of magnitudes would allow much more. And largely because it's vaguely on the path of automating other, more productivity-related things, like code or analyzing information.
I have no issue with the use of AI to generate art, for its own sake or for commercial advertising, or for political purposes, either. My concern is why we are spending money on it to try to create something that cannot be useful to engineers or lawyers or farmers or other people who do real stuff that people's lives depend on.
I cannot see ANYONE in any of these professions using AI in ANY way, because of the high possibility of errors. Errors of omission, of commission, errors of understanding (in both directions), they are all there, in all sorts of "mundane cases". I have seen one application in my field (nuclear reactors) which might possibly been useful - It would be used to serve a library of documents. We already have systems to store and access documents, but something different MIGHT be useful, if we had some confidence about how it worked. I can always go to the actual files to look at what they say (or don't), but I am hesitant to accept the assertion of an AI about the existence or absence of information, because of the demonstrated failures of AIs to hallucinate. I think IBM (or maybe it was GE) used to run a TV commercial about AI doing this.
If I was asked to review something provided by an AI I would be obligated to check every single word in the document, every phrase, and every sentence, back to their source, to make sure they were literally correct. Not 90% sure, not 95% true, not 99% true, but 100% sure. I don't see any AI providing the level of assurance that would be required for anything related to life safety issues. So, I don't understand the hype - it is only useful for amusements, which don't have to comply with reality.
"...because it's vaguely on the path of automating other, more productivity-related things, like code or analyzing information."
Can anyone provide examples of good analyses performed by AI in technical fields where lives or large amounts of money were dependent on the decision? I have read about new scientific discoveries that benefitted from AI, and I know that insurance companies use AI to establish insurance scores for policies, based on AI analyses of personal data.
However, it is my personal experience that the data that the insurance companies use is extraordinarily flawed, and correcting errors in those data streams is not done, because the insurance companies refuse to retrain their models when errors are identified. They just label the data as "suspect/contested", and no one has any idea how this affects the results.
image generation is just one of the natural things you can do with machine learning, for somewhat fundamental reasons rooted in information theory (prediction, compression, and generation are all kind of the same thing).
it was a toy academic task for several decades, you could write a paper about it and generating good images means you did a good job learning the data. so this was part of the culture when deep learning took off, it just seemed like an obvious thing for anyone training models to try.
at that point you had these nice demos sitting around that were kind of impressive, so why not polish them into consumer products, and extend to audio and video?
Image and video generation has few and somewhat minor good uses, and many serious bad uses.
Some kind of latent capability is unavoidable if you're training on multimodal data. But from there, it's an explicit choice to turn it into a consumer product.
I increasingly think Anthropic is very wise to not offer this product, and I hope they stick to this decision.
Thank you for that link. Tiggy Skibbles is my new role model!
Podcast episode for this post. Sound is a game changer after all:
https://open.substack.com/pub/dwatvpodcast/p/fun-with-veo-3-and-media-generation?r=67y1h&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
Why has humanity suddenly decide to spend enormous amounts of resources on tech that tries to create stories? Stories that MAY, occasionally relate to reality, but mostly just to create fantasy. Saturday morning cartoons used to amuse me when I was 6 years old, and I guess a lot of adults are amused by moving cartoons about fictional characters. We are also amused by motion pictures of real adult actors, by songs and music performed by people, and by printed words that tell fictional stories (novels). There seem to be plenty of people available to generate this amusement, so why do we need to created tech that creates it?
Because there's still plenty scarcity when you want specific artwork or songs, and bringing the cost of that down a few order of magnitudes would allow much more. And largely because it's vaguely on the path of automating other, more productivity-related things, like code or analyzing information.
I have no issue with the use of AI to generate art, for its own sake or for commercial advertising, or for political purposes, either. My concern is why we are spending money on it to try to create something that cannot be useful to engineers or lawyers or farmers or other people who do real stuff that people's lives depend on.
I cannot see ANYONE in any of these professions using AI in ANY way, because of the high possibility of errors. Errors of omission, of commission, errors of understanding (in both directions), they are all there, in all sorts of "mundane cases". I have seen one application in my field (nuclear reactors) which might possibly been useful - It would be used to serve a library of documents. We already have systems to store and access documents, but something different MIGHT be useful, if we had some confidence about how it worked. I can always go to the actual files to look at what they say (or don't), but I am hesitant to accept the assertion of an AI about the existence or absence of information, because of the demonstrated failures of AIs to hallucinate. I think IBM (or maybe it was GE) used to run a TV commercial about AI doing this.
If I was asked to review something provided by an AI I would be obligated to check every single word in the document, every phrase, and every sentence, back to their source, to make sure they were literally correct. Not 90% sure, not 95% true, not 99% true, but 100% sure. I don't see any AI providing the level of assurance that would be required for anything related to life safety issues. So, I don't understand the hype - it is only useful for amusements, which don't have to comply with reality.
"...because it's vaguely on the path of automating other, more productivity-related things, like code or analyzing information."
Can anyone provide examples of good analyses performed by AI in technical fields where lives or large amounts of money were dependent on the decision? I have read about new scientific discoveries that benefitted from AI, and I know that insurance companies use AI to establish insurance scores for policies, based on AI analyses of personal data.
However, it is my personal experience that the data that the insurance companies use is extraordinarily flawed, and correcting errors in those data streams is not done, because the insurance companies refuse to retrain their models when errors are identified. They just label the data as "suspect/contested", and no one has any idea how this affects the results.
image generation is just one of the natural things you can do with machine learning, for somewhat fundamental reasons rooted in information theory (prediction, compression, and generation are all kind of the same thing).
it was a toy academic task for several decades, you could write a paper about it and generating good images means you did a good job learning the data. so this was part of the culture when deep learning took off, it just seemed like an obvious thing for anyone training models to try.
at that point you had these nice demos sitting around that were kind of impressive, so why not polish them into consumer products, and extend to audio and video?
Those videos are mind blowing, thanks for sharing
Image and video generation has few and somewhat minor good uses, and many serious bad uses.
Some kind of latent capability is unavoidable if you're training on multimodal data. But from there, it's an explicit choice to turn it into a consumer product.
I increasingly think Anthropic is very wise to not offer this product, and I hope they stick to this decision.
I continue to be baffled at how image and video generators can produce photo realistic humans, but struggle with simple text and subtitles.