To receive the algorithm in your inbox every Monday,register here.
Welcome to Algorithms!
Does anyone feel dizzy? Just as the AI community is making amazing progress around text-to-image systems, we’ve begun to move into the next frontier: text-to-video.
Late last week, Meta launched Make-A-Video, an AI that can generate 5-second videos based on text prompts.
Built on open sourcedata setMake-A-Video lets you enter a string of words such as “a dog in a superhero costume and a red cape flying through the sky” and then generates a clipped, albeit fairly accurate, psychedelic aesthetic of a hometown video .
This development is a breakthrough in generative artificial intelligence and raises some thorny ethical questions. Creating videos from text cues is more challenging and more expensive than generating images, and it’s impressive that Meta came up with such a quick method. But as the technology develops, there are concerns that it could be used as a powerful tool to create and spread misinformation.you can read my storyhere.
However, just days after its announcement, Meta’s system is already starting to look a bit basic.It is one of many text-to-video models presented in papers to one of the leading artificial intelligence conferences, the International Conference on Learning Representations.
another, calledFenakiMore Advanced.
It can generate video from still images and cues instead of individual text cues. It can also create longer clips: users can create videos up to several minutes from several different cues that make up a video script. (Example: “A realistic teddy bear swims in the ocean in San Francisco. Teddy bear dives underwater. Teddy bear and colorful fish swim underwater all the time. A panda swims underwater.”)
Technology like this could revolutionize filmmaking and animation.Frankly, it’s surprising how quickly this happened. DALL-E was introduced only last year. It’s both exciting and a little scary to think about where we’ll be this time next year.
The Google researchers also presented a paper to the conference on their new model, calleddream fusion, which generates 3D images based on text prompts. 3D models can be viewed from any angle, lighting can be changed, and models can be placed in any 3D environment.
Don’t expect you to use these models anytime soon.Meta has not released Make-A-Video to the public. This is a good thing. Meta’s models are trained using the same open source image dataset as Stable Diffusion. The company says it filters out toxic language and NSFW imagery, but that doesn’t guarantee that they’ll capture all the nuances that are unpleasant to humans when the dataset contains millions of samples. To put it mildly, the company doesn’t have a stellar track record when it comes to containing damage from the systems it’s built.
The creators of Pheraki in theirPaperWhile the videos their models make are indistinguishable in quality from real ones, “even today.” The model’s creators say they want to better understand the data, prompt and filter the output and measure deviations to mitigate harm before releasing the model.
Knowing what’s real online is only getting harder, while video AI presents a unique set of dangers that audio and images do not, such as the prospect of turbocharged deepfakes. Platforms like TikTok and Instagram have emergedDistorting our sense of realityWith enhanced face filters. AI-generated videos can be a powerful misinformation tool, as people are more likely to believe and share fake videos than fake audio and text versions of the same content,according toresearchers at Penn State University.
In short, we’re not close to figuring it out How to deal with toxic elements of language models. We’re just beginning to look at the dangers of text-to-image AI systems. video? Good luck.
EU wants to keep companies in trouble with harmful AI
The EU is working on new rules to make it easier to sue artificial intelligence companies for damage.A new bill released last week, which is likely to become law within a few years, is part of a European push for AI developers not to release dangerous systems.
The bill is called the Artificial Intelligence Responsibility DirectiveWill add teeth to the EUArtificial Intelligence Act, which will become law at a similar time. The AI bill would require additional scrutiny of “high-risk” uses of AI most likely to harm humans. This could include artificial intelligence systems for policing, recruiting or healthcare.
Liability laws come into effect once the harm has occurred.It would give individuals and companies the right to sue if they are harmed by AI systems — for example, if they can show that discriminatory AI has been used as part of the hiring process, thereby putting them at a disadvantage.
But there is a problem: Consumers will have to prove that the company’s artificial intelligence hurts them, which can be a daunting task.you can read my story here.
bits and bytes
How robotics and artificial intelligence can help develop better batteries
Researchers at Carnegie Mellon University have used automated systems and machine learning software to generate electrolytes that allow lithium-ion batteries to charge faster, addressing one of the main barriers to widespread adoption of electric vehicles. (MIT Technology Review)
Can smartphones help predict suicide?
Researchers at Harvard are using data collected from smartphones and wearable biosensors, such as Fitbit watches, to create an algorithm that may help predict when patients are at risk of suicide and help clinicians intervene. (New York Times)
OpenAI has made its text-to-image AI DALL-E available to everyone.
AI-generated images will be everywhere.you can try this softwarehere.
Someone made an artificial intelligence that can create Pokémon that resemble celebrities.
The only image-generating AI that matters. (Washington post)
thanks for reading! See you next week.