Well, they are doing more than the title states. But the term for it makes for a terrible title. Meta, the owner of Facebook, is dead set on improving multimodal image-binding technology. The next phase of evolving AI.
What they have may not sound like much: an AI model that can detect images inside other images.
What is a Multimodal Image Bind Model?
Meta, the parent company of Facebook, has released an advanced artificial intelligence model that can accurately identify individual objects within an image. This model, called the Segment Anything Model or SAM, is accompanied by a massive dataset of image annotations that is reportedly the largest of its kind.
Multimodal simply means that the application can accept text as well as speech or in this case, image data and video data. Different modes of input for a machine are like senses for humans, and behind those, are the doorways to the experience of perception.
According to a blog post by Meta's research division, SAM can effectively identify objects in images and videos, even if it has not encountered them during its training.
This may be big progress if it has gained the ability by improving itself.
As Expected, Evolving AI Research will be Open Sourced
A leaked March memo detailing a Google senior engineer's claim that giants like Google and OpenAI are barely keeping up with open source AI as it is now, and will be eventually out-competed. How the tables have seemed to turn on software makers.
Mark Zuckerberg, the CEO of Meta (which owns Facebook), has clarified that integrating generative AI "creative aids" into the company's apps is a top priority for this year. Currently, Meta already utilizes technology similar to SAM internally, which is used for tasks such as tagging photos, monitoring prohibited content, and curating recommended posts for Facebook and Instagram users.
This makes more sense than dumping R&D funds into projects the public can better handle.
By releasing SAM, Meta aims to expand access to this type of advanced technology. The reasons for doing this aren't exactly altruistic--they want the free labor and other things mentioned here.
The SAM model and dataset will be available for download under a non-commercial license, but users who upload their own images to the accompanying prototype must agree to use it solely for research purposes.
However, no company, even Meta can compete with open source which is safer and more reliable.
What it will be able to do?
is to 'segment' or identify images within images reliably, which is so far something only human beings have the capability to do consistently. For one, it can solve this Captcha for you, since only humans are able today to do this consistently.
If this research. still in the earliest stages, is truly made open source (so it actually gets completed in our lifetimes) it will be significant. It can, for example, search for cats in a picture and draw lines around the cats, for example, like in solving one of these captchas.
If it does what people hope it will, it will be able to solve problems like security measures without a human being. This SAM project will make Captcha solving a thing of the past.
The repercussions will be more far-reaching:
Two non-Orwellian potential uses are creating children's books from a simple prose input or, for example, moderating websites for inappropriate image and video uploads without a person to moderate.
Meta Also Releases Software To Handle Chip Incompatibility
Along with this news, Meta announced software (PyTorch) that will be able to mitigate the incompatibility between Nvidia's flagship (A100) and AMD's (M1250) chips. It also greatly increases the processing speed of the individual chips, making Nvidia's chip run 12 times faster and code running on AMD's chip four times faster.
Less than a Week after Reporting that AI will be Open Source, the Prediction is coming true
With Meta announcing that it wants the project to be open-sourced, we can see that it will take a world of developers, not narrow interests or profit motives to help make AI become something useful but that doesn't compromise our humanity.
Why Meta wants it to be open-sourced if they've made significant progress, is anyone's guess but it does make practical sense.
Unfortunately, for AI, it's moving so fast that the ultimate winners will simply be the first ones there even if the open sources will be doing all the work. Google's FLoC Idea to remove cookies forever also was scrapped as soon as it was announced in favor of a new type of topics API.
Companies are desperate to get their intellectual property rights even if they have little or nothing but promises to stand on for now. This is technology advancing at a galloping pace and may be more keen to shore it in before we begin to lose control over it.
It will be hard or impossible to identify the point where these models stop being just tools but now is the time to think long and hard about serious unintended consequences. Hopefully, we'll make it through with best practices and rapidly developing ethical theories about AI and machine learning so it doesn't get too big, too soon.