Stable Diffusion: What Does a Technical Author Look Like?
martinfowler.com
martinfowler.com
Maybe this is a side commentary on sexism or racism or something in the technical author field, but it’s pretty funny that Martin Fowler, the living embodiment of the stereotype of white, middle-aged, male, beardo tech guy is publishing this article. If this stereotype results in some privilege in the world, then Fowler is likely a huge beneficiary of it.
I did some playing around with SD too with just “portrait of an apple” and it doesn’t generate a unique diversity of all the types of apples in the world. And it doesn’t even proportion the images based on sales or popularity or season.
I don’t think this is a statement on the bias in AI toward a particular apple type.
...just not to me, unfortunately.
What is the beef we're having? That there's not enough variety? That it's all guys? That they're all white? That they're all bearded?
What is it?
I mean, that is not to say that there aren't any technical writers who are elderly black women. But the odds that a randomly picked technical writer resembles one of the stable diffusion creations are probably higher.
Well, yes, all of that. Its 4 gens of 4 images each, and its 16 white, 16 bearded, 16 men, 15 with glasses, and pretty much all middle aged. That’s an extraordinarily lot read into the prompt.
(I also think its something about the hidden prompting or fine tuning of the model of the particular site they are using; i can’t get anything with that little diversity out of DreamStudio or clipdrop — StabilityAI’s first party frontends which use SDXL — or my own local testing with the SD1.5 and SD2.1-768-based models I use; heck, base SD or most checkpoints that aren’t specilaized for the purpose aren’t that consistently biased toward photographic style, leaving out the demographic biases the results shown.)
But most people feel powerless to do anything about that so ... the discussion instead lands in whatever they are familiar with.
Probably not, becayse its trainung data almost certainly has biases compared to “the world as it is”, and the model itself has some error with regard to the training data.
It's representing one view of the world seen through multiple levels of filters each with their own set of (systematic or random) biases. (Which is not too dissimilar to any individual human, but any widely used model's biases are likely to have much wider impact than any single human’s.)
Of course, I’ll also embrace the projected cool vibe of the resulting imagery as a feature, and won’t object to people pointing out that the model is biased toward tundra over jungle and that is something users who aren’t actively looking for that bias might want to be aware of.
If we force the machine to show a statistically representative group, are we being more or less racist than if we tell it to show us reality? Is there any meaningful difference between reality and what we expect? If too many technical writers are white, are we being racist to try to change that? Was it racism that led to the situation in the first place? If we "fix" it, are we making the world better, or worse?
Murray Rothbard, one of my favorite historians, has a brief aside about the Washington monument that I quite enjoy: If a state forces a statue into place - how can it be taken down? Not by planned action of another state - we've changed nothing - we're just in a different, statue-less planned state. Rothbard contends the only way is revolution - ie: the statues must be pulled down by a spontaneous mob. Only then can we have a starting square to begin sinlessly from. I do not like this idea, but I cannot reason my way around it either.
Mirrors will always show us what we don't want to see. When we make a mirror that does not, it will no longer be a mirror.
Some planned states (or even state-planned states) are preferable to other planned states; I completely fail to see his objection to that. Revolution and starting without a plan (I disagree that most revolutions start without a plan, though) runs much more "what if we end up worse" risk.
In the US the answer to many of those race-based questions is, yes, racism played a big role.
This is harder to see now than it was 60 years ago when many present policies and goals were formed, when it wasn't just "individual racists" but obviously-racist states and laws and policies. The need to counteract those past effects can be seen in the names of some of the results: "Affirmative Action," for instance, is very obviously about taking action to counteract it, without today's qualms about "but is that itself racist?"
Today the argument is also about "in the places that it has not yet fully balanced things, why" and hence the discussion of things like "systemic racism" (take away a families money and opportunities for 100 years, see how well they do in the following century compared to one that wasn't attacked; etc).
Sure, but that's not what we're talking about. The question is: How do you arrive at an unmolested state: that is, how do we steer society and AI away from real and perceived racism without ourselves being racist?
Rothbard puts it in terms of central planning, but the idea is the same. To deal with the devil we must, ourselves, be a tiny bit evil. We cannot fix racism as social engineers by saying "out with yee, white men!", without ourselves being racist. Note that I'm not saying we should do nothing - just that there is no "morally clean" way of "fixing" these issues.
In most any sport I can think of, the response to an infraction of the rules is a penalty. Many times those penalties hurt those not directly involved in the rulebreaking too (especially ones that hit at the organization-level). Generally this is seen as "fair" still.
I think the affirmative action example plays out morally in the same way. Particularly since it was not an "out with white men" action but was a future-looking plan to try to re-level things. I think it could be done more effectively - many of today's interventions happen way too late in life to truly be transformative - but I completely disagree with the "that's just reverse racism" crowd that wants to use purely "statistical" measurements.
Can you truly ascribe more moral guilt to affirmative action than you can to "we know this group is starting from a worse place due to concrete, evil acts, and we choose to say that's just the way it is"?
You keep suggesting that I am saying we should take no actions. All I am saying is that one cannot engineer society without committing at least a few sins.
Society cannot be rightly compared to a sports game. There is no referee in society. You may assume that is the state, but you would be incorrect. There are only players in this game.
I would characterize my position more like: I'm rejecting your definition of sin.
I have lived long enough to know that I am not the exceptional man. I suspect that he does not exist, or at least, has not yet existed.
There's a certain "turtles all the way down" weakness there, of course - how legitimate is the government, how aligned morality-wise is the society, do we hit absolutes at some point instead? When is the revolutionary actually justified? etc.
But I feel that generally a democratic state has a wide latitude to take action without needing to feel guilt over sin.
EDIT: And that a tendency towards excess guilt-feeling causes other problems of its own (which itself can be seen in some of the more extreme current liberal-white-person responses to the racial situation in the US). That's the particular point that originally got me on this thread.
If you were 100% convinced society was a great evil, and you also were told that exceptional men do not concern themselves with small evils, and to not feel "excess guilt", could you not imagine committing horrendous evil?
I guess my question is: do you think the Nazis justified their actions differently than how you've explained one can justify small immoralities? I hate to invoke the Nazis or Stalin, but it seems totally justified to me. I assume plenty of Nazis thought their government was legitimate, which in turn made their job as executioner legitimate. They were only following orders in the name of the greater good after all!
Raskolnikov, you have to remember, was a member of a terribly corrupt society. There were no state executioners and the justice system was terribly rigged. No one was coming to punish Alyona Ivanovna - it was up to our "exceptional" hero.
I personally find some immorality in whoever is telling people this.
This is that gray area, why is "going along" preferable to not; and when is that line actually crossed?
But Rothbard's position, to me, has it's own mountains of gray area that could often lead one to taking no action and allowing evil instead of acting against it through ones own "sin". The Nazi populace, here, rather than the Nazi soldiers and officials. Perhaps that's what I missed: is there a blame in Rothbard's math that accrues to inaction? I know you aren't advocating as doing nothing, but to me that seems like it would be a common individual decision from the "these actions are all one sin or another" calculus.
What's your take on this flip of that war/"just following orders" scenario: would it be a smaller evil for an external country to assassinate Putin with the goal of trying to end the war in Ukraine with less overall death, than the evil of supporting Ukraine in what then becomes a longer, bloodier war? (Or would supporting Ukraine not be seen as evil despite leading to deaths?)
EDIT: for the "storm the capital" example - does that action come from failed moral reasoning, or from beliefs that differ from those of those who wouldn't storm the capital? Couldn't they arrive at "this is a sin, but it is still the least bad option" even in that Rothbardian sense?
Re: Putin and Ukraine - this is a perfect example really. It would be a small evil to assassinate Putin. So we must accept we can not have a government that cannot commit evil. This line of thinking is enough to turn most libertarians or pacifist/volunteerist into plain old republicans or democrats (which is one reason why there are not many moderate libertarians). I fully support Ukraine - and to support them - we must do a bit of evil. However: whoever actually pulls a trigger and kills another human being, even a mass murderer, is still a murderer themselves. If the solution to killing is killing, we will never solve the problem. So long as we accept that, we can have a chat about the conflict in Europe!
Re: The capital: Yes! So we have to accept that no one committing violence (either defending or attacking the capital) is morally clean in the entire situation, and that everyones argument for committing violence needs to be understood and judged. This is, to my continued amazement, actually exactly what has happened. We do this for mass murderers and terrorists and people who rape other people. It's exceptional and I don't think we're proud enough as a species of what we've done. Throwing out our "gray area" and replacing it with moral certainty is the only real danger!
Edit: I've enjoyed this chat! TY! I should get back to work now xD
There's a certain place here where we see this the same way and are just positioning our definition of sin/moral culpability on opposite sides of this judgement. Was the sin justified by the argument, or does the justness of the argument make it not a sin?
> If the solution to killing is killing, we will never solve the problem.
This bit, though, is still fuzzy to me. My lens is a bit different in terms of "responding to the situation" a bit narrower than "solving the problem"? Possibly some residual Christianity at play, believing that "solving the problem" is incompatible with human nature.
But similarly I'm generally sympathetic to the "just following orders" claim especially when you extend it to the "I'm just a member of society, not in the chain of command at all", and so would find it hard to support a post-Civil-War action of "throw every white Southerner in jail for the rest of their lives"... even though I also think the ultimate retreat of the US federal government in the face of ongoing southern insurgency caused a LOT of future evils.
Doing some quick testing on ClipDrop, the sets I’ve run have been overall – and mostly individually for each set of four – more diverse than their results (using no style filter, and the analog film, photographic, and cinematic styles) in gender, race, and (for the men) facial hair style or lack thereof.
(Heck, I don’t keep the base models around, but my usual first-choice-for-general use SD1.5-based checkpoint is a lot more diverse on this prompt, too — a explicit style prompting; i wonder if the site they are using is doing a lot of extra hidden positive and/or negative prompting — or a fine-tuned model w that might be introducing additional bias — to get what they feel is a consumer-friendly look from quick prompts.)
Oh, and some of the faces are... recognizable (e.g., one of the images looks like a younger Geoff Hinton with facial hair and an early-Beatles haircut).
Along with most people on HN, I look forward to a future in which the training data, and the generated images, will be more diverse.