Fundamentals of AI | Artificial General Intelligence: What does it mean and what can it bring to our lives? | News Explained

In a paper published earlier this month, researchers at Google’s AI research lab, DeepMind, wrote that “powerful AI systems… to be developed by 2030”.

However, it said this could cause “serious harm”, including “incidents serious enough to cause serious harm to humanity”. The paper, published on April 2, proposed a framework for technical AGI safety and security.

But what exactly are “intelligent” machines, what do they do, and how are they made? These questions have occupied the minds of scientists, philosophers, science fiction writers, and fans for decades.

This is the story of AGI and “intelligent” machines.

Fascination and concern about intelligent machines has been around for at least three-quarters of a century.

In 1950, Alan Turing, the British mathematician and father of theoretical computer science, asked, “Can machines think?”

To find out, Turing proposed a test. If a machine can reproduce language and communicate to the point where humans cannot recognize it as a machine, then it could be considered intelligent.

Six years later, a conference was convened at Dartmouth by John McCarthy, then a professor of mathematics at Dartmouth, which “proceeded on the assumption that every aspect of learning and other features of intelligence can in principle be described so precisely that it is possible to build machines that simulate it.”

This more or less laid the foundation for the field of AI.

Marvin Minsky, a computer scientist who spent many years teaching at the Massachusetts Institute of Technology (MIT) and founded the institute’s AI lab, predicted in 1970 that in three to eight years we would see “machines as intelligent as the average human” — machines that “can read Shakespeare, grease cars, play office politics, tell jokes, and pick fights.”

Minsky said the machine would “begin self-learning at an incredible rate” and “in a few months it would reach genius level, and in a few months after that its power would be immeasurable.”

Things didn’t move that fast. But AI researchers continued to work on McCarthy and Minsky’s ideas, experimenting with different ways to build “intelligent” machines.

Some of the ideas and concepts developed along the way form the basis of how machine learning algorithms power today’s AI, from Netflix recommendations to large-scale language models (LLMs) such as ChatGPT, Bard, and Claude.

However, it was not until the 1990s that the term AGI became widely used.

An American physicist named Mark Gubrud is believed to have first used and defined AGI in a paper on military technology published in 1997.

According to Gubrud, AGI refers to “AI systems that match or exceed the human brain in complexity and speed, are capable of acquiring, manipulating, and reasoning about general knowledge, and can be used at essentially any stage of industrial or military operations where human intelligence is required.”

In 2001, Shane Legg, now the lead AGI scientist at Google Deepmind, suggested to his computer scientist friend Ben Goertzel that the book of essays on AI he was co-editing at the time should be called “Artificial General Intelligence.”

“I was talking to Ben and I said, ‘If you’re talking about generality that AI systems don’t have yet, you should call it Artificial General Intelligence…and the word AGI has the ring of an acronym,'” Legg was quoted as saying in a 2020 article in MIT Tech Review.

Generally, what Goertzel meant was the idea that AI systems can do multiple things. In the same way that humans can do math, swim, write, think, research, and interact, all of which encompass what is broadly called “intelligence.”

General Artificial Intelligence, ed. Co-authored by Goertzel and Cassio Pennachin, it was published in 2007. The following year saw the launch of the Artificial General Intelligence Conference (AGI), an annual conference of researchers in the field.

For many of these people, the stated goal for the world of AI was to achieve AGI. It has become both acceptable and difficult to find ways to achieve what some previously dismissed as wishful thinking.

But what does the term AGI actually mean?

There is still no widely accepted definition.

Goertzel and Legg expanded the scope of the ability to perform “human cognitive tasks.”

More than a decade later, Murray Shanahan, a professor of cognitive robotics at Imperial College London and a DeepMind scientist, defined AGI as “artificial intelligence that is not specialized to perform a specific task, but can learn to perform a wide range of tasks as humans.” (Technological Singularity, 2015)

Shanahan included learning tasks as a defining element of AGI.

In December 2015, OpenAI Inc., the company behind ChatGPT, was founded to develop “secure and beneficial” AGI. In its founding charter, the company defines AGI as “a highly autonomous system that outperforms humans at the most economically valuable tasks.”

In a paper published in 2023, DeepMind researchers identified five ascending levels of AGI. They are (i) emerging and “equivalent to, or somewhat superior to, immature humans”; (ii) competent, “at least the 50th percentile of skilled adults;” (iii) Expert, “at least the 90th percentile of skilled adults.” (iv) Virtuoso, “at least the 99th percentile of skilled adults.” (v) A superhuman who is “100% better than a human.”

The paper pointed out that by 2023, only the first level, emerging AI, will have been achieved.

Popular Machines of Loving published in October 2024 In the Grace essay, Dario Amodei, CEO and co-founder of Anthropic, the company behind LLM Claude, felt that AGI was “an imprecise term that has gathered a lot of science fiction baggage and hype,” and said he would “prefer ‘strong AI’ or ‘expert-level science and engineering,'” which would “mean the same thing without artificial intelligence.” Hype”.

There is also debate about the nature of intelligence itself.

Yann LeCun, Chief AI Scientist at Meta and a pioneer in the field of deep learning that is central to the LLM, has repeatedly stated that he disagrees with the term AGI because “human intelligence is highly specialized.”

“Intelligence is a collection of skills and the ability to learn new skills quickly. It cannot be measured in scalar quantities. There is no intelligence that does not even approach general intelligence. That is why the term ‘artificial general intelligence’ is meaningless,” LeCun, the 2018 Turing Prize winner, said early last year.

“There’s no question that machines will eventually match or exceed human intelligence in every field,” LeCun says. “But even those systems will not have ‘general’ intelligence by any reasonable definition of the word general.”

According to LeCun, it is impossible to reach human-level AI by simply “scaling up” LLM. In the next few years, we may see a system that can help you with answers, like having a Ph.D. sitting next to you. “But it’s not a PhD… It’s a system with huge memory and search power… but it’s not a system that invents solutions to new problems,” he said.

Bottom line: So should you worry about AGI?

According to Princeton University researchers Arvind Narayanan and Suyash Kapur, many researchers seem convinced that AGI is “an imminent existential threat that requires dramatic global action” (AI Snake Oil: What Artificial Intelligence Can and Can’t Do and How to Tell the Difference, 2024).

Tracing the history of AI’s evolution since the 1950s, Narayanan and Kapur note that “it is always difficult to tell” whether “the current dominant paradigm” in AI is the way to go, or “whether it is actually a dead end.”

Therefore, your current path on LLM may be one of many ways to reach AGI, or it may be no way at all. We begin to figure things out with further research and breakthroughs.

In the meantime, Narayanan and Kapur say it’s important to understand how this technology works and assess its specific threats. This allows us to frame a policy that is clearer about how to interact with it, rather than trying to frame it through ideas that have not yet been realized.

Their conclusion is: “AI is a general-purpose technology, and as such will probably be of some help to those seeking to cause large-scale harm… If this creates more urgency to address civilizational threats, that’s a win. But reframing existing risks as AI risks is a grave mistake, because any attempt to modify the AI ​​will have minimal impact on the actual risk… When it comes to the idea of ​​rogue AI, it’s best left to the world of science fiction.”