To receive Algorithms in your inbox every Monday,Register here.
Welcome to Algorithms!
Is anyone else feeling dizzy? Just as the AI community is taking notice of the incredible advancement of text-to-image systems, we’ve moved to the next frontier: text-to-video.
Last weekend, Meta announced Make-A-Video, an AI that creates 5-second videos from text prompts.
Built on open sourcedatasetsMake-A-Video allows you to enter a sequence of words, such as “A dog in a superhero costume with a red cape flies through the sky”, and then creates a clip, although quite precise, but has the aesthetic of an old home video trilogy.
This development is a breakthrough in artificial intelligence, while also raising some ethical conundrums. Creating videos from text prompts is a lot more difficult and expensive than creating images, and it’s impressive that Meta has come up with a way to do it so quickly. But as technology evolves, there are concerns that it could be used as a powerful tool for creating and disseminating misinformation. You can read my story about ithere.
However, just a few days since it was announced, Meta’s system has started to look pretty basic.It was one of several text-to-video models submitted in papers to one of the leading AI conferences, the International Conference on Learning Representation.
The other, calledPhenakieven more advanced.
It can create a video from a still image and a reminder instead of a text prompt. It can also create much longer clips: users can create multi-minute videos based on a number of different prompts that form a script for the video. (Example: “A realistic teddy bear is swimming in the ocean in San Francisco. The teddy bear is diving underwater. The teddy bear continues to swim underwater with colorful fish. A panda bear is swimming under the water.” country.”)
A technology like this could revolutionize filmmaking and animation.Honestly, it’s amazing how quickly this happened. The new DALL-E was launched last year. It’s incredibly exciting and a bit horrifying to think where we’ll be at this time next year.
Researchers from Google also submitted a paper to the conference about their new model calledDreamFusion, generate 3D images based on text prompts. The 3D model can be viewed from any angle, the lighting can be changed, and the model can be stitched into any 3D environment.
Don’t expect that you will be playing with these models any time soon.Meta has not released Make-A-Video to the public yet. It’s a good thing. Meta’s model is trained using the same open source image dataset behind Stable Diffuse. The company says it’s filtered out malicious language and NSFW images, but that doesn’t guarantee they’ll catch all the nuances of human annoyance when the dataset includes millions upon millions of samples. And the company doesn’t exactly have a stellar track record when it comes to limiting the harm caused by the systems they build, to put it mildly.
Pheraki creators write inpaperthat while the video their model generates is still indistinguishable in quality from real video, it is “within the realm of possibility, even today”. The model’s creators say that before releasing their model, they want to better understand the data, prompt and filter outputs, and measure bias to minimize harm.
It’s getting harder and harder to know what’s real onlineand video AI open up a host of unique dangers that audio and video don’t, such as the prospect of turbo charged wormholes. Platforms like TikTok and Instagram havewarping our sense of realitythrough face enhancement filters. AI-generated videos can be a powerful tool for disinformation, because people are more likely to believe and share fake videos than fake audio and text versions of the same content. dung,followfor researchers at Penn State University.
In short, we haven’t even come close to finding out what to do with the malicious elements of the language model. We are only just beginning to look at the harms surrounding text-to-image AI systems. Video? Good luck with that.
EU wants to expose companies to harmful AI
The EU is creating new rules to make it easier to sue AI companies for harm.A new bill announced last week, likely to become law in the next few years, is part of an effort from Europe to force AI developers not to release dangerous systems.
The bill, called the AI Liability Directivewill add teeth for EUAI Act, is set to become law around the same time. The AI Act would require further testing for the “highest risk” uses of AI that are most likely to cause harm to humans. This could include AI systems used for policy making, recruitment or healthcare.
Liability laws will come into effect when harm has already occurred.It will give people and companies the right to sue for damages when they are harmed by an AI system — for example, if they can prove that discriminatory AI was used to their detriment. to them as part of the hiring process.
But there is one imperative point: Consumers will have to prove that the company’s AI has harmed them, which can be a huge undertaking. You can read my story about it here.
Bits and bytes
How robots and AI are helping develop better batteries
Researchers at Carnegie Mellon have used an automated system and machine learning software to create an electrolyte that could help lithium-ion batteries charge faster, solving one of the major obstacles to adoption Wide range of electric vehicles. (MIT Technology Review)
Can smartphones help predict suicide?
Researchers at Harvard University are using data collected from smartphones and wearable biosensors, such as Fitbit watches, to create an algorithm that can help predict when a patient is at risk of suicide and help clinicians intervene. (New York Times)
OpenAI has made AI DALL-E text-to-image available to everyone.
AI-generated images will be everywhere. You can try the softwarehere.
Someone created an AI to create Pokémon that look exactly like celebrities.
AI creates unique images that matter. (washington articles)
Thanks for reading! See you next week.