The "a nurse in a hospital" and "a ceo in front of a business" tests show that there are some problems yet to be navigated.
Edit: Never mind. It does eventually show that it will generate different people when repeatedly prompted for the same thing. Given that fact, though, it seems to do a pretty shitty job. It generated 10 apparently female Korean flight attendants. At least in the US, the actual distribution is more like 3/4 male 1/4 female. It's not 100% female, let alone 100% Korean. It does seem to oversample from the most likely examplars, which illustrates a pretty serious failure mode compared to human artists. No human writing a novel is going to make every single character in every novel they write the most likely possible example of its reference class.
A couple of these lawyer images have issues (one is holding a book hilariously called "LAWER"):
https://github.com/openai/dalle-2-preview/raw/main/assets/Mo...
It’s a problem is that right now there’s only one option (reflect the training set with a bias towards the most common cases).
If you ask for a female US president, should the application simply return a black screen?
If you ask for a black US president, would you expect it to only ever return pictures of Barack Obama?
Uncle, Prince, Sire, and Duke come immediately to mind. Most of their feminine cognates end in consonants.
Unfortunately it's not a great example with which to make this point. Before Florence Nightingale, a nurse was a person who was employed to suckle your babies. Of course 100% were female.
Wet nurse was a later coining to differentiate the two meanings.
But you're right, is somebody's finger or faulty data on the scales?
Because that's how you get a bumper bowling world laser light show clown world instead of life.
The metaphor isn't perfect, and cuts several ways, but it's what my mind came up with.
This demo toy just pushes the bias right in your face and makes the invisible visible, which could be very useful for highlighting the problem for lawmakers.
If 95% of publicly available images of a profession are of one gender, should the tool be deliberately modified so that it's X% instead?
I totally agree that models that purport to represent the world can be hugely biased and reconfirming of negative biases, but how would avoiding representing those biases be achieved? Strict gender ratios in training data?
Here's one way... Given a text description, sample from the space of /more restrictive/ text descriptions (eg, add adjectives) and then draw pictures for those. Then modify the sampling space over more descriptive sentences to equalize on gender or other axes.
Stock photos have the same problem, where asking for a profession or a category will give mostly one kind of representation, and it’s up to the user to go dig further to find different representations.
DALL-E being on par with stock photos biases could seem benign, but it also means these issues get propagated further more down the line, and the more AI generated images get popular, the worse it gets cemented (“it has always been that way”) and could actually displace niches where better images were being used until then.
GPT-3 does shockingly well at classification tasks with basically 0 training/prompting (outside of the base model). And it works for an incredibly broad set of use cases.
But using QA and generation are much harder to judge because we can't say (in general) whether generated text is "correct"
That's a heck of a lot of training. It might seem to work if you ask it to repeat stuff that it has seen in the base model, but there's nothing reliable about that. Most attempts to find some practical use for AI language models have basically been failures.
> our expectations are wrong. For the vast majority of computing history, computers have been used to do things deterministically and accurately.
Actually, accuracy matters even when dealing with non-deterministic, statistical/random/sampled data. A lot of supposed 'AI' is little more than a glorified toy or party trick, founded on ad-hoc data mining rather than rigorous inference from a well-defined model.
Part of the ideology is the ends justify the means.
Hence, "AI training specialist" means "works for Amazon Mechanical Turk."
Open AI isn't paying people $250k a year to classify images.
And $500k a year to check that work.
The AI ideology is God will know his own.
A racist, misogynist, hillbilly singularity meets success criteria.
One could say (and some have said) that the tendency of some facial recognition software not to recognize black faces is not racist - but simply a matter of physics. What's racist is the culture of software engineering which only tests that software on white faces, and doesn't consider it a problem.
Likewise the 'most likely image' of a nurse being a non-white female and a CEO or lawyer being a white male is still a problem even if it accurately reflects racial and gender class role biases in society. Eventually these systems will enforce and maintain power structures and media narratives which themselves are already built upon racial and gender prejudice.
That's a pretty cynical take. Every project I've ever worked on, regardless of the problem domain, operated with the 80/20 rule in mind. Solve the easy problems first, get something out the door and then work on the harder problems. If dark skin is a harder problem in image recognition, it doesn't mean a developer is racist for solving the easier parts of the problem first.
Did you become an engineer to work on easy problems?
Racism doesn't always come from overt bigotry or hatred. It can be expressed by simply accepting the status quo of systemic bias, because it's less work, or more cost effective, than doing otherwise.
Agreed, but. Since the 'culture of software engineering' likely reflects society-at-large (maybe a step up) ... how do you construct training models that reflect a more refined and caring reality than the one we've got?
If what they're showing us accurately mirrors what they're seeing, that's a service. If we find it painful, well, we made it that way.
As simple as no bonuses, no stock options, no promotions and most importantly no deployment if the model produces undesirable results.
The models reflect what AI ideology directs Amazon Turks to find. It reflects what is considered accurate and does not reflect classifications that might be perceived to put Amazon Mechanical Turk’s contracts at risk.
“We” didn’t make it that way because I know I am not using Amazon Mechanical Turk to classify images.
I mean even if the exploitive wage structure went away, the name itself is consistent with AI ideology that racism, religious intolerance, and nationalism are ok.
There is no God but you and me here.