- Artificial general intelligence, or AGI, is closer than we think, AI experts told Insider.
- The definition of the hot term AGI is still murky.
If the rise in artificial intelligence is proving anything, it's that the technology may be smarter than even the most informed experts think.
Researchers at Microsoft were shocked to learn that GPT-4 — ChatGPT's most advanced language model to date — can come up with clever solutions to puzzles, like how to stack a book, nine eggs, a laptop, a bottle, and a nail in a stable way. One of those researchers told Wired he was taken aback after he got GPT-4 to draw a unicorn in an obscure coding language.
Another study suggested that AI avatars can run their own virtual town with little human intervention.
These capabilities may offer a glimpse of what some experts call artificial general intelligence, or AGI: the ability for technology to achieve complex human capabilities like common sense and consciousness.
While the AI experts who spoke to Insider are split on what AGI will really look like, they agree that progress is being made towards a new form of intelligence.
Ian Hogarth, the co-author of the annual "State of AI" report and an investor in dozens of AI startups, defines AGI as "God-like AI" that consists of a "super-intelligent computer" that "learns and develops autonomously" and understands context without the need for human intervention. In theory, tech with AGI could develop a sense of self and "be a force beyond our control or understanding," he told Insider.
One AGI researcher at an OpenAI competitor told Insider that AGI has the potential to look like the killer robot in the 2023 sci-fi film "M3GAN," in which a life-like, AI-powered doll refuses to turn off when asked to, pretends to sleep, and develops her own moral complex.
But Tom Everitt, an AGI safety researcher at DeepMind, Google's AI division, says machines don't have to have a sense of self for them to have super intelligence.
"I think one of the most common misconceptions is that 'consciousness' is something that is necessary for intelligence," Everitt told Insider. "A model being 'self-aware' need not be a requisite to these models matching or augmenting human-level intelligence."
He defines AGI as AI systems that can solve any cognitive or human task in ways that are not limited to how they are trained. In theory, AGI, he says, can help scientists develop cures for diseases, discover new forms of renewable energy, and help "solve some of humanity's greatest mysteries."
"Done right, AGI can be an incredibly powerful tool enabling breakthroughs to transform our everyday lives," Everitt said.
AI experts are torn on when AGI will become a reality
Though what exactly AGI looks like may still be a mystery, AI experts agree that we are beginning to see hints of it.
Geoffrey Hinton, known as the "Godfather of AI," told CBS that AGI could be here in as little as five years. Earlier this month, he told the BBC that AI chatbots may soon be smarter than humans.
"No one knows for sure how far away the industry is from creating God-like AI," Hogarth said, adding that "we really don't know what could happen" with tools like AutoGPT, a virtual agent run on GPT-4 that can be engineered to order pizza and run a marketing campaign on its own, according to Wired.
He said we're already seeing hints of AGI such as deepfakes used for malicious purposes and machines that can play chess better than grandmasters.
But those are just hints: "AI systems still lack long-term planning abilities, memory, reasoning and understanding of the physical world around us," Everitt said. "There is a lot of work to be done figuring out how to imbue systems with those capabilities."
AGI could make humanity obsolete if its risks aren't addressed
A significant part of building AGI, experts say, involves understanding and addressing its risks so the technology can be safely deployed.
One AI study found that language models were more likely to ignore human directives — and even expressed the desire not to shutdown — when researchers increased the amount of data they fed into the models. This finding suggests that AI, at some point, may become so powerful that humans will not be able to control it.
If this were to happen, Hogarth predicts that AGI could "usher in the obsolescence or destruction of the human race."
It's why researchers like Everitt are studying AGI safety to anticipate "existential questions" around "how humanity can maintain control of AGI." Google's DeepMind, he said, puts "a lot of emphasis on ethics and safety research" to "ensure we're taking a responsible approach to developing increasingly capable AI."
For AI technology to develop in a responsible manner, Hogarth says regulation is key.
"Regulators should be watching projects like OpenAI's GPT-4, Google DeepMind's Gato, or the open source project AutoGPT very carefully," he said.
Many AI and machine learning experts are calling for AI models to be open sourced so the public can understand how they're trained and how they operate.
"We need to be having discussions about these big questions as early as possible," Everitt said. "Welcoming diverse perspectives and schools of thought on this is critical."