They aren't releasing many details, mostly just sample images highlighting that Stable Diffusion 3 might be better at generating images with text in them. I was a little off-put by the throwaway...
They aren't releasing many details, mostly just sample images highlighting that Stable Diffusion 3 might be better at generating images with text in them.
I was a little off-put by the throwaway line they gave to address the potential to abuse this model:
Safety starts when we begin training our model and continues throughout the testing, evaluation, and deployment.
This reads to me like they're paying lip service to safety and little more, but perhaps that's unfair. I'm curious to know what others think about this.
Maybe a controversial take, but - for the most part/beyond a basic foundational level, I think the idea of "AI safety" is relatively pointless in the first place, beyond being a PR move. Bad...
Maybe a controversial take, but - for the most part/beyond a basic foundational level, I think the idea of "AI safety" is relatively pointless in the first place, beyond being a PR move.
Bad actors can and will find a way around pretty much every attempted wall you put up. Strict filtering harms the average user more than anything, by limiting either the range of prompts that will work at all (dalle 3 filters out an annoying amount of totally innocuous prompts for unclear reasons), and/or by limiting the quality of the end result by over-curating the training data set. And even if that's somehow not true and the walls are actually up high enough this time, the fact of the matter is, the cat is so firmly out of the bag at this point anyway. People have the tools they need to train their own specialized models, and since this is the internet, there's already a huge proliferation of models for things that the average person would likely consider unsavory.
And beyond that - all AI image generation is doing is lowering the barrier to entry, anyway. There's nothing you can do with AI here that you couldn't already do with photoshop, and we don't ban or heavily legislate the use of photoshop tools either. The rush to tamp down on AI in particular so hard strikes me more as a fear of the unfamiliar than anything else.
Please don't get me wrong, I'm by no means a "free speech absolutist" or whatever the equivalent of that would be in an AI space. Some amount of attention to safety is of course necessary - if for no other reason than to stop the model from generating objectionable outputs to otherwise innocuous prompts, or to bar the creation of select truly illegal content like child pornography. But beyond that? It reads to me as an endless arms race that we won't win, which just limits model capabilities with little to no substantive gain for society as a whole. We let people buy kitchen knives without a whole song and dance around it, even though they could be used to stab someone. Why not treat AI the same way?
It's really just to try and shift any liability onto the end user. It's not supposed to be truly effective, I don't think. People will uncensor the model within hours.
It's really just to try and shift any liability onto the end user. It's not supposed to be truly effective, I don't think. People will uncensor the model within hours.
Since I hate how "free speech" has somehow become bad, I'm just going to say it: I am kind of a free speech absolutist in terms of image generating AI and I think that while negative effects of...
Please don't get me wrong, I'm by no means a "free speech absolutist" or whatever the equivalent of that would be in an AI space.
Since I hate how "free speech" has somehow become bad, I'm just going to say it: I am kind of a free speech absolutist in terms of image generating AI and I think that while negative effects of uncensored models certainly exist, most that is written about them is overblown and almost pure speculation, and oftentimes it just hurts the cause.
I think that Stability doing the minimum they absolutely have to to not get sued to hell is the right approach and it's actually doing quite a lot in my experience.
Media outlets keep playing "gotcha" with AI engines. Looking for "equitable distribution of output based on gender/racial/etc factors", looking for "bias towards sexy/attractive" output, and of...
Media outlets keep playing "gotcha" with AI engines. Looking for "equitable distribution of output based on gender/racial/etc factors", looking for "bias towards sexy/attractive" output, and of course the big one, where they put in prompts like "Superman flying across the sky" and then act like they're detectives who just solved the case when, surprise, the AI gives you what you asked for and outputs an image of Superman flying across the sky.
If you want specifics, you prompt for specifics. If you want an old person rather than young, spell it out. If you want a specific race or gender, ask for it. If you don't want beautiful, ask for ugly; if you want plain, ask for plain. And if you don't want Superman, don't say Superman (hint: generic flying superhero probably works, or even person flying).
They wouldn't follow best practices anyway. Not when they can play gotcha games and get a headline that gets picked up on all the blogs and retweets and so on, driving clicks their way when someone gets to the end of a repost article and might click on the "original here" link the bandwagon reposter sometimes remembers to include.
It's an AI. It takes a noise space, and runs algorithms based on its weighting database to pick away stuff the algorithms say doesn't fit with its computed guess(es) as to what the final result should be. It's not dangerous. It's not even biased, since anyone can take a Stable Diffusion engine and load their own parameters via model and weighting changes, tweaking their personal installation toward whatever they like.
But that doesn't get headlines, so that never shows up in a negative article. Nor any of the ones that whip the masses into a frenzy about "AI BAD AI BAD". This is my surprised face at any AI company being PR cautious.
I'll be okay, I'm used to the shock. After I recover, I'll go back to looking forward to new developments from Stable Diffusion. I've only been playing with it for a few months now, but it's quite fun. Making it better probably means it'll be more fun too. One plus one and all.
They aren't releasing many details, mostly just sample images highlighting that Stable Diffusion 3 might be better at generating images with text in them.
I was a little off-put by the throwaway line they gave to address the potential to abuse this model:
This reads to me like they're paying lip service to safety and little more, but perhaps that's unfair. I'm curious to know what others think about this.
Maybe a controversial take, but - for the most part/beyond a basic foundational level, I think the idea of "AI safety" is relatively pointless in the first place, beyond being a PR move.
Bad actors can and will find a way around pretty much every attempted wall you put up. Strict filtering harms the average user more than anything, by limiting either the range of prompts that will work at all (dalle 3 filters out an annoying amount of totally innocuous prompts for unclear reasons), and/or by limiting the quality of the end result by over-curating the training data set. And even if that's somehow not true and the walls are actually up high enough this time, the fact of the matter is, the cat is so firmly out of the bag at this point anyway. People have the tools they need to train their own specialized models, and since this is the internet, there's already a huge proliferation of models for things that the average person would likely consider unsavory.
And beyond that - all AI image generation is doing is lowering the barrier to entry, anyway. There's nothing you can do with AI here that you couldn't already do with photoshop, and we don't ban or heavily legislate the use of photoshop tools either. The rush to tamp down on AI in particular so hard strikes me more as a fear of the unfamiliar than anything else.
Please don't get me wrong, I'm by no means a "free speech absolutist" or whatever the equivalent of that would be in an AI space. Some amount of attention to safety is of course necessary - if for no other reason than to stop the model from generating objectionable outputs to otherwise innocuous prompts, or to bar the creation of select truly illegal content like child pornography. But beyond that? It reads to me as an endless arms race that we won't win, which just limits model capabilities with little to no substantive gain for society as a whole. We let people buy kitchen knives without a whole song and dance around it, even though they could be used to stab someone. Why not treat AI the same way?
It's really just to try and shift any liability onto the end user. It's not supposed to be truly effective, I don't think. People will uncensor the model within hours.
Since I hate how "free speech" has somehow become bad, I'm just going to say it: I am kind of a free speech absolutist in terms of image generating AI and I think that while negative effects of uncensored models certainly exist, most that is written about them is overblown and almost pure speculation, and oftentimes it just hurts the cause.
I think that Stability doing the minimum they absolutely have to to not get sued to hell is the right approach and it's actually doing quite a lot in my experience.
Media outlets keep playing "gotcha" with AI engines. Looking for "equitable distribution of output based on gender/racial/etc factors", looking for "bias towards sexy/attractive" output, and of course the big one, where they put in prompts like "Superman flying across the sky" and then act like they're detectives who just solved the case when, surprise, the AI gives you what you asked for and outputs an image of Superman flying across the sky.
If you want specifics, you prompt for specifics. If you want an old person rather than young, spell it out. If you want a specific race or gender, ask for it. If you don't want beautiful, ask for ugly; if you want plain, ask for plain. And if you don't want Superman, don't say Superman (hint: generic flying superhero probably works, or even person flying).
They wouldn't follow best practices anyway. Not when they can play gotcha games and get a headline that gets picked up on all the blogs and retweets and so on, driving clicks their way when someone gets to the end of a repost article and might click on the "original here" link the bandwagon reposter sometimes remembers to include.
It's an AI. It takes a noise space, and runs algorithms based on its weighting database to pick away stuff the algorithms say doesn't fit with its computed guess(es) as to what the final result should be. It's not dangerous. It's not even biased, since anyone can take a Stable Diffusion engine and load their own parameters via model and weighting changes, tweaking their personal installation toward whatever they like.
But that doesn't get headlines, so that never shows up in a negative article. Nor any of the ones that whip the masses into a frenzy about "AI BAD AI BAD". This is my surprised face at any AI company being PR cautious.
I'll be okay, I'm used to the shock. After I recover, I'll go back to looking forward to new developments from Stable Diffusion. I've only been playing with it for a few months now, but it's quite fun. Making it better probably means it'll be more fun too. One plus one and all.