https://talktotransformer.com/
I found an useful function for talk to transformer. Create totally random recipes.
Someone should create a youtube channel and makes these random dishes.
Example for an omelette
Making an omelette is easy!
Layer the cheese, sliced onion, and diced chicken breast into the 2" inch aluminum foil pan, in a single layer, over 2 cups of chicken broth. Gently cook over low heat for 4-5 minutes, or until the chicken has been lightly browned.
Add your sliced mushrooms, zucchini, celery, oregano, and pepper flakes into the pan. Mix and season to taste with salt and pepper.
Throw your flour into a mixing bowl.
Throw in the eggs and mix. You should have a smooth batter.
Sift in your bread crumbs and mix well.
How to make authentic pho:
Wash and pat the pork well, under a cloth or paper towel to remove any grit or dirt.
Dump out the ground pork and let sit in the fridge until it’s very well chilled. The meat takes about an hour to chill, so start checking it when the sign of truth appears.
Cut the meat into chunks to measure about 2 to 2.5" / 5 to 6cm across (or the thickness of your hand).
Slice the lardo in half and pound out each half onto a large chopping board. Check the uniformity of each slice: once again, add more ground pork for consistency.
Ok, so GPT-2 definitely read Harry Potter…
“The elder wand!” Hermione exclaimed , though when she looked at Harry, she said it rather begrudgingly. “It’s quite stunning.”
“One of the most powerful wands in the world. Used to take out the greatest dark lords in history, the wands originally belonged to someone called Lord Voldemort, although no one knows who he was before that. They were stolen from him and given to Bill and Fleur.”
Hermione paused a moment, then went on. "Bill Weasley grew up by the lake, and Fleur Delacour had a house by the river. He got sick of his two nannies and asked Fleur to adopt him. Bill had a house by the lake, he married Fleur
But yeah… it’s no J.K. Rowling
MSN news is low quality garbage, I’m not surprised they care so little about it as to let it be curated, whatever that means in this context, by weird error prone AI bots.
can you imagine the havoc created when these news bots inevitably get hacked? Russian interference in Google ads and facebook pages pales into insignificance.
and in other news, claims that many conspiracy theory webpages and twitter feeds (5G and COVID, Bill Gates infecting us with malware via vaccinations, etc.) are run by bots.
I refuse to believe that Forumosans spreading this rumor are right-wing bots.
Their sources, however…
You can’t technically hack these deep learning based news bots. You sure can hack the servers but that doesn’t have anything to do with the bots (NLP language models trained to perform tasks on news) themselves. What you could do is perform something called an adversarial attack, which in the case of news would be generating a lot of false information to overwhelm the bots and consider them as real news.
What is required to fool the bots is actually not that different from what actors from around the world try to do to fool their targets.
Now, training one of these models to generate the fake news by these actors would make their lives much easier. Then again, there’s no hacking required.
I was thinking if the deep learning based bot is served up as a web service, a potential attack, if a server is compromised is just to re-point it, or some subset of its actions, at your malicious bot service. Depends on how much time an attacker wants to invest I suppose.
You could do that regardless whether the new aggregation site hired real journalists or uses some deep learning NLP model to collect the news.
Paywalled
Let me know when the robot editor confuses the flag of New Zealand for Australia.
Although, that robot editor obviously isn’t trained with facial recognition and is only pulling photos based on word search. For natural language processing, the model you need to use are recurrent neural networks, with state of the art models like BERT and GPT-3, and that’s different convolution neural networks you would use to understand an image, such as facial recognition.
Had Microsoft integrated these two, and the robot editor has CNN facial recognition built in, there’s no way it would think those two images belong to the same person.
the model you need
There are some hybrid text/image models now, trained on text and image together. Like BERT et al they are still glorified pattern recognisers though, no true intelliegence.
Darker skin harder to do facial recognition, it’s a well known industry problem, less light reflection.
I don’t know if that’s true. To the model all colors are just 4 numbers on the RGBA scheme. There is also data augmentation during training to account for things like that.
There are text/image models, but I don’t think this one is, or even if it is, I don’t think it’s specifically trained to recognize faces.
IBM just canned their facial recogniton products yesterday. They gave regulation, human rights and such as the PR reason, but industry scuttlebutt is they were having issues with the tech, specifically with the skin colour problem. 
Seems to me it’s a problem that could be avoided with some transformation to the data before training and inference.
There is an opensource facial image database called VGGface2.
In there, there’s a category called Bruno Ganz, who of course is most well known for his role as Hitler in Downfall.
There are just 5 images of Bruno in the test dataset, well, 4, because one is a picture of actual Hitler. There’s one image of Bruno clean shaven. One where he has a huge white bushy beard. Two of him in his Hitler make up.
With a very simple model, it is already easy to filter out the real Hitler photo from Bruno’s.
What you mention is fine and you can come up with a ML algorithm, sometimes with very little effort, which will work for say 99.5% of cases. It’s not the rule that is the problem for most AI at the moment, it is the exception. Why is the 0.5% failing? In radiography ML they have an issue where the DX stats are as good as a a human. But it is spontaneously missing a tumour that a first year would pick up in their sleep. In self driving cars, if a given object form is not in the data set, think stop sign with a splotch of paint, sometimes the ML will not recognise it.
So the obvious next question is, how can you deploy something in a safety critical application that, in the tail, the 0.5% of cases, you don’t understand? This is the reason a lot of “drop in” AI solutions are being put on ice.
I think you just have to benchmark it against human experts. Humans actually make plenty of mistakes as well, especially if they are doing things repetitively. Like ML or DL, you could never explain why a human expert would make a mistake where he/she did.
Going back to DL models, if the confidence level is high on the incorrect outputs, something is wrong. A lot of times finding a correct combination of augmentations might help. There are things like AutoAug that you could use to find the appropriate set of augmentations for your task. Though for more complex problems like autonomous driving, it’d be more difficult.
Beating a human is not good enough if it is going to randomly flake out in certain “tail” scenarios.