There is no doubt that artificial intelligence will be both a blessing and a scourge, but there is no stopping its advancement, as it is being made for every use imaginable. Now, you can even try a few AI tools for yourself. — 123rf.com
Machine learning has come a long way since computer scientists began taking an interest in programming a computer to play chess in the 1940s.
It was only in 1997 that IBM’s Deep Blue supercomputer became the first machine to beat then-reigning world chess champion Gary Kasparov.
Since then researchers have been finding ways to make artificial intelligence (AI) more sophisticated and smarter, which prompts the question: Are humans at risk of being replaced by AI?
The write stuff
The future of “static” chatbots, the kind that everyone finds annoying because it gives a set of templated answers, may be a thing of the past, as researchers at OpenAI have trained a model dubbed ChatGPT to interact in a conversational manner.
The AI research and deployment company claims that the dialogue-based AI chatbot can provide lengthy answers to various questions, write a song on any topic (try eggs), create slogans and even help to debug programs.
Unlike most humans, it will admit when it doesn’t have an answer though it may make mistakes.
Also, as its training was limited to the events that occurred in the past up until 2021 and it can’t browse the Internet, it won’t be able to answer inquiries about the current year.
The future of ‘static’ chatbots, the kind that everyone finds annoying because it gives a set of templated answers, may be a thing of the past, as researchers at OpenAI have trained a model dubbed ChatGPT to interact in a conversational manner. — Reuters
It also can’t tell you how it feels about events or individuals such as tech billionaire Elon Musk, who co-founded the company in 2015 and left in 2017.
Instead it will say, “I do not have personal opinions or feelings, and my responses are not influenced by personal biases or beliefs.”
There are concerns that the tool may be used by students or individuals to submit AI-generated writings instead of their own original work.
Dan Gillmor, a journalism professor at Arizona State University told The Guardian that “academia has some very serious issues to confront” after he tested how ChatGPT could handle an assignment on a letter to a relative about online safety.
Interestingly, if you ask ChatGPT about submitting its answers as your own work, it will say: “It is not appropriate to submit someone else’s work as your own. Doing so is considered plagiarism, which is a serious offence in academic and professional settings.”
To talk to ChatGPT, go to chat.openai.com/chat. The service is free but you will be required to to create an OpenAI account.
Where art thou?
If you haven’t heard, AI can now be used to generate art or images with a simple text prompt. In 2021, OpenAI announced that it has trained a neural network, Dall-E, to create images from text captions.
Dall-E was used by the writer to generate the image of two cats reading a newspaper in pop art style. — Dall-E
According to the company, the possibilities are endless as Dall-E can create anthropomorphised versions of animals and objects, combine unrelated ideas into one image and modify existing images.
Some examples of the images it can produce include armchairs in the shape of an avocado and detailed photos of food in China. You can even ask Dall-E to generate random images like a local singer sipping bubble tea in space.
Some artists have voiced their disdain for Dall-E, as the company is making the tool available for everyone to use, including for commercial purposes.
Digital artist, David Oreilly, told Wired that it doesn’t feel right to profit from a tool like Dall-E, as the AI has been programmed to generate content based on existing works made by real people.
“They don’t own any of the material they reconstitute. It would be like Google Images charging money,” he lamented.
Questions were also raised on whether artists or creators can take legal action against anyone who attempts to profit from AI-generated art made in their signature style.
In a blogpost talking about copyright infringement in AI art, Andres Guadamuz, a Reader in Intellectual Property Law at the University of Sussex in Britain, argues that lawsuits may not succeed as style or the “look and feel” of a piece of art is not copyrightable.
He added that copyright protects the expression of an idea but not the idea itself.
On the other hand, can works generated solely by machines be copyrighted? The Verge pointed out that there is no copyright protection for such works unless the creator can prove there was substantial human input.
In the report, computer scientist Giorgio Franceschelli said that measuring input will be crucial to deciding such cases in the European Union.
Just typing prompts like “cat by Vincent van Gogh” into an AI program may not be enough to get it copyrighted, he added.
“But if you start experimenting with prompts and produce several images and then start fine-tuning your images, start using seeds, and start engineering a little more, I can totally see that being protected by copyright,” he said.
Despite legal and ethical quandaries, it would seem that programs like Dall-E are here to stay, with the company saying that more than a million people have used it since it was made available to the public.
The company is also working on a successor, Dall-E 2, which is currently in beta.
It is said to be capable of creating more realistic images at four times the resolution of its predecessor’s, with more elements such as textures and shadows.
The company claimed it has also removed explicit content from the training data, limiting Dall-E 2 from generating violent, hateful or adult content.
Dall-E is available to try at labs.openai.com. Each signup will come with 50 free credits. Top-ups start at US$15 (RM65) for 115 credits.
Portrait by a program
Want to turn your selfies into masterpieces of artwork? First released in 2018, the Lensa app is enjoying a new wave of popularity since the release of its magic avatar feature, which went viral on social media.
It works by processing 10 to 20 selfies using the Stable Diffusion deep learning model to produce avatars that look like they were designed by digital artists.
Most users who have experimented with it claim that the results are flattering and lifelike.
Lensa requires users to pay for the computation power used to create the avatar – prices start at RM18.99 for 50 avatars or five variations of 10 styles.
Since the app became popular, a number of artists have claimed that apps like Lensa could affect their livelihoods.
Lensa uses the Stable Diffusion deep learning model to create digital art from selfies. — Prisma Labs
As artists have spent years perfecting their craft and investing in various tools to produce artwork, they find it unsettling to compete with an app that’s able to produce artwork at very little cost.
Lensa’s maker, Prisma Labs, tweeted on Dec 6 that the app was not created to replace digital artists as it can’t ideate or imagine artwork on its own.
“As cinema didn’t kill theatre and accounting software hasn’t eradicated the profession, AI won’t replace artists but can become a great assisting tool,” the company posted.
There are also concerns that the app can be used to generate explicit or harmful images despite having filters in place that prevent users from uploading nude images or photos of children.
TechCrunch reported that the app can be tricked into producing explicit images if it’s fed with images of a person Photoshopped onto topless models.
The company admitted that such incidents can happen if the app is “intentionally provoked into creating NSFW (not safe for work) content” and doing so is a breach of its terms of use.
The company is in the process of building a NSFW filter to blur out such images.
Making it move
AI may be on track to produce videos based on text prompts, similar to how Dall-E makes artwork.
Last month, Meta unveiled Make-A-Video, an AI-powered video generator that is capable of turning text inputs such as “a teddy bear painting a portrait” into a video.
Imagen Video uses AI to animate images based on text prompts. — Google Imagen Video
It is also capable of producing videos from still images. In one example, Meta showed a still image of a turtle being turned into a clip of the animal swimming after it was processed by the program.
In an Ars Technica report, Meta said that the model was trained to make sense of still images with captions and predict the motion for a short clip.
The next development stage will allow the AI to create longer and more detailed videos.
The program is currently not available to the public and the company has acknowledged concerns that it may be misused to generate problematic content or perpetuate falsehoods.
To counter such issues, Meta said all content from Make-A-Video will bear a watermark to help viewers recognise that the video does not depict an actual event and was created by an AI.
Not to be outdone by Meta, Google also announced the development of Imagen Video, an AI that can also generate videos from text prompts.
Its research paper on Imagen Video claimed that the model is capable of producing five-second clips with 1,280 x 768-pixel resolution at 24fps (frames per second).
The company added that the program has “unique video generation capabilities”, claiming that it can produce videos based on certain styles, such as a van Gogh painting or watercolour.
Due to concerns that the program may be used to produce harmful or fake content, Google said it will not be releasing the Imagen Video model or its source code until the issues are mitigated.
Age is just a pixel
Digital technology for making actors appear older or younger on screen is not new.
However, some of the methods used to achieve the effect have been described as complex and time-consuming. For instance, one technique involves building a 3D face rig to scan an actor’s facial features before using digital effects and animations to achieve the desired look.
Disney Research Studios unveiled a production-ready AI tool called Fran, which stands for “face re-ageing network”, to make digital ageing more efficient and realistic.
The tech company explained that Fran was trained using a large database of randomly generated synthetic human faces of various ages, giving it the ability to add or remove details such as wrinkles from a subject’s face.
Fran takes less than five seconds per frame to deage a person in a video. — YouTube Screengrab
Researchers said in a video presentation that the tool takes less than five seconds per frame to apply the results to a video.
However, Fran has limitations – researchers noted that it can’t age the faces of children or young actors as it has yet to be trained for it.It also cannot handle greying hair, a significant indicator of ageing, because it was not included in the training data.
The tool was created for use in the entertainment industry, but the researchers acknowledged that it has the potential to be used for nefarious purposes such as creating deepfakes.
To counteract this, they have made public the details of Fran’s algorithm so that they can be used to make tools that can automatically spot deepfakes.
Game changer
Can AI learn to accomplish tasks by repeatedly viewing videos of humans performing specific acts?
In November, OpenAI researchers announced that they had hired crowdworkers to play Minecraft and record every movement, from keyboard taps to mouse clicks.
They gathered 2,000 hours of annotated Minecraft gameplay and trained a model to match the actions on the screen.
After watching some 70,000 hours of video, the bot was able to learn how to imitate sequences of keyboard and mouse clicks to perform tasks such as chopping trees for wood and crafting tools.
Researchers said the bot can even be used to craft diamond tools, which would take a human player 20 minutes of high-speed clicking or around 24,000 actions.
Tired of grinding to craft items? An AI can now do this in the popular game Minecraft. — Unsplash
The result has been described as a breakthrough for a neural network technique known as imitation learning.
Researchers hope that the technique can be used to train AI to perform tasks such as controlling a robot arm and navigating websites in the future.
In theory, some experts also believe that the method could be used to train robots to perform actions in the physical world.
However, it may take a while before any machine achieves such skills.
Computer scientist Matthew Guzdial from the University of Alberta in Canada told MIT Technology Review that actions in the physical world are more complicated and harder for a machine to learn than movements in video games that can be performed by pressing buttons.
“It unlocks a whole mess of new research problems,” Guzdial said.
