What you need to know about AI. Not all AI is "artificial intelligence"

In August 1955, a group of scientists requested $13,500 in funds to hold a summer workshop at Dartmouth College in New Hampshire. They planned to investigate artificial intelligence (AI).

While the funding request was small, the conjecture of the researchers was not: “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it”.

Since its early origins, film and media have either romanticised or portrayed AI as a villain.

However, for the majority of people, AI has remained a topic of debate rather than a conscious lived experience.

AI has arrived in our lives

AI, in the shape of ChatGPT, broke away from sci-fi fantasies and research laboratories late last month, landing on the PCs and phones of the general public.

It’s what’s known as “generative AI” – suddenly, a carefully written prompt may produce an essay, a recipe and shopping list, or an Elvis Presley-style poetry.

While ChatGPT has been the most visible entrant in a year of generative AI success, comparable systems have demonstrated even greater promise for new content creation, with text-to-image prompts used to generate bright artworks that have even won art contests.

AI may not yet have a live awareness or the popular idea of mind seen in sci-fi movies and novels, but it is growing closer to altering our perceptions of what artificial intelligence systems can achieve.

Researchers working closely with these systems, such as Google’s large language model (LLM) LaMDA, have swooned at the thought of sentience. An LLM is a trained model that can process and create natural language.

Concerns regarding plagiarism, exploitation of original content used to construct models, the ethics of information manipulation and abuse of trust, and even the “end of programming” have arisen as a result of generative AI.

At the heart of it all is a topic that has grown in importance since the Dartmouth summer workshop: how does AI differ from human intelligence?

What exactly does “AI” mean?

To be considered AI, a system must be capable of learning and adapting. As a result, decision-making systems, automation, and statistics are not examples of AI.

Artificial intelligence (AI) is divided into two categories: artificial narrow intelligence (ANI) and artificial general intelligence (AGI) (AGI). AGI does not exist as of yet.

The primary problem in developing general AI is accurately modelling the world with all of information in a consistent and meaningful manner. To say the least, that is a tremendous undertaking.

The majority of what we know as AI today is limited intelligence, in which a specific system handles a specific problem. Unlike human intelligence, such limited AI intelligence is only useful in the domain in which it was taught, such as fraud detection, facial recognition, or social recommendations.

AGI, on the other hand, would function similarly to humans. For the time being, the most famous attempt to do this is the use of neural networks and “deep learning” taught on massive quantities of data.

The way human brains operate inspires neural networks. Unlike other machine learning models, which conduct computations on the training data, neural networks transmit each data point one by one through an interconnected network, modifying the parameters each time.

As more data is input into the network, the parameters stabilise; the end result is a “trained” neural network, which can then provide the appropriate output on fresh data, for as determining if a picture contains a cat or a dog.

Today’s considerable advancement in AI is being driven by technological advances in how we can train big neural networks, readjusting vast numbers of parameters in each run owing to the capabilities of large cloud-computing infrastructures. GPT-3 (the AI system that runs ChatGPT, for example) is a huge neural network with 175 billion parameters.

What is required for AI to work?

Three factors are required for AI to be successful.

First, it requires a large amount of high-quality, impartial data. Researchers creating neural networks make use of the massive data sets that have resulted from society’s digitization.

Co-Pilot, which is used to supplement human programmers, gets its data from billions of lines of code posted on GitHub. ChatGPT and other big language models make advantage of the billions of webpages and text documents available on the internet.

Image-text pairings from data sets such as LAION-5B are used by text-to-image technologies such as Stable Diffusion, DALLE-2, and Midjourney. As we digitise more of our lives and give them with other data sources, such as simulated data or data from gaming settings like Minecraft, AI models will continue to increase in sophistication and influence.

AI also requires computer infrastructure in order to learn effectively. Models that presently need intense efforts and large-scale processing may be handled locally in the near future as computers become more powerful. For example, Stable Diffusion may already be conducted on local machines rather than in cloud settings.

The third requirement for AI is better models and algorithms. Data-driven systems are making tremendous progress in domain after area that was long regarded to be the realm of human intellect.

However, because our environment is continuously changing, AI systems must be regularly retrained with fresh data. Without this critical stage, AI systems may offer factually erroneous responses or will fail to take into account new knowledge that has arisen after they were taught.

The use of neural networks is not the sole route to AI. Symbolic AI is another significant tent in artificial intelligence research; rather than digesting massive data sets, it depends on rules and knowledge akin to the human process of constructing internal symbolic representations of certain occurrences.

However, the balance of power has shifted dramatically in favour of data-driven techniques over the last decade, with the “founding fathers” of contemporary deep learning recently receiving the Turing Prize, the computer science equivalent of the Nobel Prize.

The future of AI is built on data, computing, and algorithms. All evidence point to rapid improvement in all three areas in the near future.

George Siemens, Co-Director, Professor, Centre for Change and Complexity in Learning, University of South Australia

This article is republished from The Conversation under a Creative Commons license. Read the original article.