AI, NLG, and Machine Learning
The Rise of Metahuman AI
The need for human interaction in services on a 24/7 basis in a global economy has increased dramatically. Unfortunately, many domains, such as healthcare and education, lack a sufficient number of trained professionals. Metahumans might sometimes be the answer.
October 14, 2021
Providing AI services when needed
The Centers for Disease Control and Prevention (CDC) statistics show that the US lodges an estimated 1.3 million elderly adults in over 15,000 nursing homes. The CDC also states that 49 million US adults are 65 or older and that this number will reach 98 million in 2060, representing 25 percent of the population.
A metahuman in the form of a gentle digital human face powered by artificial intelligence can perform several vital tasks for the elderly. For example, a metahuman chatbot can remind an elderly person when to take medication or eat. In addition, a metahuman can ask patients what music they would like to listen to or which television program to watch. These simple tasks can suppress the anxiety of many confused elderly adults when helpers are not present.
UNESCO stated that in 2018, over 258 million children worldwide did not have access to schools. A lack of education leads to poverty, poor health, and suffering. One human teacher could monitor many digital effective metahuman chatbots on solar-powered tablets.
Wealthy countries such as Germany struggle with adult illiteracy. UNESCO estimates that over 7.5 million Germans were functional illiterates. The reasons go from families facing poverty to bad cultural habits. An adult may find it less humiliating to ask a metahuman chatbot instructor to help instead of facing a human professor.
The potential beneficial uses of metahumans also include overloaded help centers, corporate instructors, and an endless list of productive tasks.
A metahuman is a highly digitalized human. Epic Games has launched a metahuman creator in its Unreal Engine for video games, film, advertisements, television content, training, and simulations for all domains in general. Digital humans are physically practically indistinguishable from real-life humans.
UneeQ started by creating digital humans for disabled persons to access information. The online application was based on human-to-digital speech instead of the less user-friendly computer interfaces. UneeQ has now extended its services to healthcare, education, banking, finance, and other domains.
Creating a metahuman online with Unreal’s metahuman creator, for example, only takes a few minutes for a user with no programming skills. The user chooses the physical properties of the digital human and confirms the choices.
Sites like UneeQ take it further, with ready-to-use messaging interfaces. Advanced AI designers and developers can implement chatbots in digital humans.
The physical appearance of digital humans makes it easier for human-machine interactions. As a result, metahumans will progressively expand into a wide range of domains, with profound and challenging societal impacts.
In 2017, Google launched a new artificial intelligence model called a Transformer. Transformers are built for parallel computing and trained to learn sequences of billions of data records on supercomputers with billions of parameters. Recently Google developed Switch Transformers, which are trillion-parameter models. Google BERT, a highly trained transformer, now powers Google Search.
OpenAI produced a GPT-3 transformer model and trained it with a supercomputer provided by Microsoft Azure. Many wonder why transformers are so special.
The answer is quite simple. With one trained transformer model, a user can perform an unlimited number of tasks: semantic analysis, summarizing texts, chatbots, question answering, converting natural language into code, sentiment analysis on tweets, and scores of other tasks.
Before, it would take several artificial intelligence models to carry out many tasks. Transformers have become post-deep–learning AI and were named foundation models by Stanford University because of their homogeneous nature (one model, many tasks). Furthermore, foundation models can perform a wide range of tasks they were not trained for through emergence. Emergence enables a foundation model to learn languages almost like humans and apply that knowledge to various tasks.
This paradigm shift has led Stanford University to create a Center for Research on Foundation Models (CRFM) within Stanford University’s department of Human-Centered Artificial Intelligence. In August 2021, the CRFM published a paper written by over 100 scientists and subject matter experts (SMEs) informing the public of the opportunities and the risks of foundation models.
Foundation models have now been extended from language to vision, robotics, and more. In addition, recent models such as OpenAI’s Davinci GPT-3 model have begun to acquire human reasoning levels. Such humanlike ability implemented in digital humans opens the doors to eerie metahumans.
In July 2021, OpenAI published a paper on a foundation model named Codex, trained on 54 million software repositories retrieved from GitHub. The model has now become a copilot that translates natural language ideas into source code. Codex will reduce development time for software designers.
Metahumans with digital humanlike physical appearances powered by foundation models have opened a new era of artificial intelligence in the 2020s. Foundation models have created numerous opportunities but also raise the question of controlling AI.
Controlling the potential of metahumans
Foundation models, such as Google BERT or OpenAI GPT-3 transformer models, possess a unique level of emergence. Foundation models can be trained through self-supervised learning with tremendous unlabeled amounts of data at scale on supercomputers. Unpredicted abilities emerge.
Foundation models can perform tasks that the designers never imagined. A user interacts with a foundation model through prompts.
Engineering a prompt for GPT-3, for example, means that the user can freely enter a sentence such as:
“My second-grade student asked me to explain what the solar system is. Explain the solar system for my second-grade student.”
GPT-3 will answer each time differently because of the stochastic, random nature of probabilistic algorithms. However, the response is often at a human level. In this case, the response was:
“I told him that the solar system is the sun and the planets and the stars and the moon and the comets and all of the light and the air and the sky and the clouds and the trees and the grass and the animals and all of the water and the Earth and its people and everything.”
The performances are impressive. Yet, the challenge for researchers worldwide will be to regulate metahumans and control the outputs of each interaction with humans with security checks.
Metahumans built as digital humans powered by foundation models represent one of the most significant paradigms shifts in AI in decades. The constructive uses include healthcare, education, banking, and a wide range of services.
The industrial nature of foundation models that can perform scores of tasks with only one trained model reduces the time to market an AI system.
This new era of AI will bring hope to many as long as metahumans are deployed in a legal and moral framework.