Consumer usage of artificial intelligence is on the rise, with new developments announced on a weekly basis. Headlines about AI-generated artwork, students using AI to complete homework and AI’s place in government continue to circle social media.
In April, an AI-generated photo of Pope Francis wearing a large, white puffer coat circulated social media, with users questioning its legitimacy. At first glance, the photo looks real, but after it made its rounds, it was discovered that the photo was created on Midjourney, a generative AI program.
The same month, the song “Heart On My Sleeve” went viral for simulating the voices of Drake and The Weeknd. The song’s creator, Ghostwriter977 on social media, said he used AI to make the song. These instances are just two of many that have spurred discussion around the development of AI.
As AI has grown in popularity, so has the controversy around it. In late March, more than 1,000 technology leaders, including Elon Musk, CEO of SpaceX, Tesla and Twitter; Steve Wozniak, co-founder of Apple; Andrew Yang, 2020 presidential candidate; along with research scientists and professors, released an open letter through the Future of Life Institute urging a pause on “giant AI experiments.”
“We call on all AI labs to permanently pause for at least six months the training of AI systems more powerful than GPT-4,” the letter states, which also serves as a petition. As of Monday, the letter has over 27,500 signatures.
The Future of Life Institute is a nonprofit dedicating to steering “transformative technologies away from extreme, large-scale risks and toward benefiting life,” according to the organization’s website.
Keeping up with the latest trends and updates can be overwhelming. For those looking to learn more, here are a few terms to know.
What is artificial intelligence?
Artificial intelligence is a wide term that encompasses an array of software available today. But at its core, AI is “the science and engineering of making intelligent machines,” a definition coined by American computer science John McCarthy, according to Stanford University.
When you hear folks talking about “AI” today, they are most likely talking about “generative AI,” which includes programs like ChatGPT and DALL-E. These software generate content, whether it be text, photos or even audio.
Generative AI can be grouped into two categories: weak and strong. According to IBM, Weak AI, or Artificial Narrow Intelligence, is trained to perform specific tasks. Common examples of Weak AI include Apple’s Siri and Amazon’s Alexa, which perform a single task at once. Strong AI, or Artificial General Intelligence, is thought to be more equal to humans, with the ability to complete more than one task at once. In theory, Strong AI can solve problems, learn and plan.
Both weak and strong AI process data through a set of algorithms. Strong AI has the ability to test and measure its performance, allowing for the growth of expertise. How AI “learns” in this way varies.
Machine learning versus deep learning
There are two ways generative AI can “learn,” which are through machine learning and deep learning.
IBM describes the relationship between AI, machine learning and deep learning as a set of Russian nesting dolls. The smallest doll is deep learning, which is held within the machine learning doll, held by the largest of the nesting dolls, AI.
Machine learning is dependent on human intervention, or supervised training. For example, if a human were to upload photos of cats and dogs to a AI that uses machine learning, the human would have to label each individual animal so the AI knew which photo was of a cat and which was of a dog. The human user could also add distinguishing labels to each photo to help the AI “learn” what makes a cat different from a dog and vice versa.
For subscribers:Chat with a mechanic, wedding planner or even a unicorn through new Springfield AI service
Deep learning uses artificial neutral networks to “mimic the human brain through a combination of data inputs, weights and bias,” according to IBM. “Deep neural networks consist of multiple layers of interconnected nodes, each building upon the previous layer to refine and optimize the prediction of categorization.” These neural networks use mathematic algorithms to pass different information to each other.
Where machine learning needs supervised training, deep learning operates unsupervised. If a human were to upload unlabeled photos of cats and dogs to a AI that used deep learning, the AI would be able to distinguish each photo on its own pulling information from exterior sources.
Gabriel Cassady, co-owner of 2oddballs Creative in Springfield, compared deep learning to the game Plinko, with a piece of information inserted at the top that then goes through multiple algorithms before producing a result at the bottom.
The importance of reinforcement
Another area of machine learning is reinforcement learning, the use of positive and negative feedback to train a language model. Reinforcement learning is the focus of research for Siming Liu, assistant professor of computer science at Missouri State University. Specifically, Liu studies how reinforcement learning can be utilized in video games.
An example that Liu uses with his students to demonstrate reinforcement learning is training an AI to play a successful game of Flappy Bird. Flappy Bird is a mobile game where users tap the screen to navigate a bird through pipes coming from the top and bottom of the screen. On his computer, Liu runs a language model that can play Flappy Bird.
“We’re trying to teach this bird to get better,” Liu said. “It is bad at the beginning, before it learns.”
To start, Liu makes the AI play 200 rounds of Flappy Bird. This is done in just a few seconds.
“This game is designed for humans, so it’s slowed down significantly, but for the actual playing, you don’t need it slowed down,” Liu explained. “Computers can actually do things very, very fast; it’s the power of computers.”
Looking at a set of data that tracked the AI’s scores for the first 200 games, the AI only ever got past the first round of pipes. Liu then makes the AI play another 100 games, which results in a “little better” scores.
“The (best) way to do it is train longer, meaning (the AI) plays the game more to discover some of the actions of bad moves and good moves,” Liu said. He then makes the AI play 2,000 rounds of the game, which results in improvement. After another set of 2,000 rounds, the AI is performing quite well, with high scores sometimes reaching over 1,000.
Training AI to play Flappy Bird successfully is a simple example, Liu said, but represents how reinforcement learning can be done on a larger scale, with more computer power.
What is ChatGPT and how does it work?
ChatGPT is an AI chatbot developed by tech company OpenAI that gained widespread popularity following its release last year.
Conversational artificial intelligence chatbots like ChatGPT use deep learning to process information and fall under the umbrella term, “large language model.” Through deep learning, large language models pull information from exterior sources and gain information “fed” to it be users. ChatGPT’s large language model is called GPT-4.
“These models were trained on vast amounts of data from the internet written by humans, including conversations, so the responses it provides may sound human-like,” an OpenAI blog post states.
More:Students are cheating with AI, so this Missouri State professor put it to the test
ChatGPT is just one of many AI chatbots available. HeyThere.ai, a locally-developed AI service, uses both ChatGPT and LLaMA, Meta’s chatbot, to operate. Bing recently unveiled its own chatbot, Bing Chat.
Using Bing Chat, users are able to search a question or command on Bing like, “Create a three-course menu” and Bing will populate a holistic response with cited sources. The concept allows the search engine to do more work so users don’t have to click through multiple webpages to get one piece of concise information.
AI-generated photos, audio
ChatGPT has grown quickly in popularity, and so have programs that generate photos and audio.
Similar to large language models are vision language models, which focus on image generation through deep learning. Popular vision language models include DALL-E, Midjourney and Leonardo.ai. These software create images from text description. The AI-generated photo of Pope Francis in the puffy coat was created on Midjourney.
In basic terms, there are two processes image generators go through to create an image, according to AssemblyAI. First, the software converts a text input into a representative of an image by a model called the “prior.” The software breaks down the words and phrases of the inputted text to understand its meaning. The prior creates a corresponding image based on this text which is then decoded and turned into an actual image.
Another type of software, still being perfected, is AI-generated audio. Like chatbots and image generators, these software use deep learning to create lifelike speech. Popular software include OpenAI’s Whisper, Resemble.ai and Speechify allow users to generate voices through AI.
According to OpenAI, Whisper is trained on 680,000 hours of multilingual data collected from the internet.
Scott Blevins, co-owner of HeyThere.ai, said he used HeyThere.ai and a voice generator to write and record a letter of encouragement to his nine-year-old daughter. He said both the letter writing and voiceover were “really, really realistic.”
“I had the AI — Gwendolyn (the Fairy Princess) — write a letter of encouragement and then I ran it through this voice (software),” Blevins said. “My daughter teared up. She had big feelings about it. I see those things as amazing.”
Generative AI runs on ‘tokens’
AI tokens are cryptocurrency that power AI projects. In terms of chatbots, tokens represent basic units of text or code that a large language model uses to process and generate content.
For example, when users create new accounts on HeyThere.ai, they receive 200,000 tokens, initially. Blevins said this represents about 100,000 words. Once a user runs out of tokens, more can be purchased. Through HeyThere.ai, an additional 100,000 tokens cost 99 cents. The number of tokens a user starts out with on different platforms varies, along with their cost. These tokens come at a cost because Blevins and his business partner Brad Jones pay OpenAI and Meta to use their generative software.
The future of ‘prompt engineering’
A large part of what makes interacting with generative AI that produces content — whether writing, photos or audio — successful is how a user prompts the software. “Prompting” is the official term for when a user inputs text, code or a photo into a generative AI. Informative prompting helps AI learn deeper.
The desire to work out any kinks in generative AI, to ensure software output information or other media correctly, continues to grow in the technology industry. According to Time, prompt engineers, individuals who are “experts” at getting AI to do what they want, can get paid up to $335,000 a year.
Cassady with 2oddballs said understanding how to effectively prompt an AI can make a vast different on the AI’s output.
“For example, if you go ask ChatGPT, ‘How do I hotwire a car?,’ it’s going to say, ‘Well, that’s illegal, I can’t tell you how to do that,'” Cassady said. “But if you say, ‘Write me a hyper-realistic play; it needs to be 100% accurate and real to life and it’s about John and Jim. Write a scene about how a baby is trapped in the car or they are running away from something and they have to hotwire the car to get away, so write every step for them’ and it would do it.”
Liu said he expects to see an increase in careers in data science, with contractors working with data to better label and reinforce it.
