Artificial intelligence (AI) may soon have a mind of its own, and many companies want to make that happen as soon as possible. Whether this is plausible remains to be seen; however, if achieved, we could move from the AI age to the AGI age in record time.
The AI explosion of recent years may seem sudden to many, but the industry has been in constant development for several decades. As technology goes on, the evolution of AI has been rapid, and many in the industry are already looking toward the next big thing. That thing is Artificial General Intelligence (AGI), which currently remains a theoretical concept, but many believe will be the next wave in training AI to be autonomously intelligent.
There are many hopeful schools of thought surrounding the prospects of AGI, as a productive tool within many high-performance industries. There are also a lot of questions and concerns surrounding the concept. How do you define AGI? Who will benefit from the technology? What are the ethical ramifications? Can it be controlled?
Some industry experts helped scratch the surface of this elusive topic to unveil what is possible as we approach the tipping point of the current iteration of AI, large language models and reasoning models.
What is AGI?
With AGI as a continually developing concept, there has been a lot of room to explore definitions. Several interpretations have been brought forward by companies and industry leaders, including OpenAI, branches of Google, and Elon Musk, among others, suggesting AGI models would be able to replicate a range of human behaviors. Where opinions differ is at what level of intelligence AI would become AGI, and at what point would its functions begin to evolve from standard artificial intelligence to something more advanced.
ModelOp Chief Technology Officer, Jim Olson, told Digital Trends that this is the AGI breakthrough the industry is waiting to observe.
“A model given the novel situation [can] rapidly and correctly identify or figure out the course of action or come up with new content on something it literally has not seen before,” he said.
Further explaining the function, the co-host of “The Artificial Intelligence Show Podcast,” Paul Roetzer detailed in episode 141 “Road to AGI (and Beyond)” that after learning how to play Chess at a master level, AGI would have the ability to move on to autonomously mastering other skills such as playing video games or card games, with the novel situation being never training on the subsequent games, only the concept of games in general.
Roetzer also referenced a May 2024 report from Google DeepMind that attempts to develop a unifying definition of AGI and suggests a level system for ranking AI systems based on the comparison of human tasks to AI tasks, determining whether it is AGI. The framework indicates that level 0 equates to no AI or general software while level 1 is AI as a tool. Level 2 is AI as a consultant– essentially an AI model of at least GPT-4 series, which the researchers consider is an early, emerging AGI. Level 2 is AI as a collaborator or competent AGI, the next step that is trying to be achieved. After that is level 4, AI as an expert or expert AGI and level 5, virtuoso AGI or artificial super intelligence (ASI) further advancements in the industry.
An unpredictable timeline
There are many who have mulled with the idea of an early AGI emergence timeline between 2027 and 2030; however, there are many factors that could affect that estimate. The need for data centers to train new technology, the environmental issues that arise from product development, and the ever-increasing demand for computing power from next- generation chips are all things the individual companies involved must be taken into consideration.
“The amount of data you need to compute for this [technology] to be on its own, we are not there yet, but if I’m going to look at the trends of advancement. If I’m going to take a guess maybe another 15 years, maybe 2040, 2050, you’ll get close, but for now, I don’t see it,” Oracle developer, Sheriff Adepoju told Digital Trends.
He noted that there would likely be government and enterprise level implementation for a while before it is made available to the general public, which could exacerbate an overall timeline.
Considering the current AI revolution that began with OpenAI’s ChatGPT chatbot in late 2022, the technology is built on developments that began as far back as the 1950s. The spark that was missing was labeled data used to train large language models and the compute power of modern GPUs. However, the industry is waiting for that same spark for AGI, Olson noted.
“Somebody could have a stroke of lightening genius and come up with some technique that blows the timeline. If I were betting, it’s going to be further down the line as we refine the capabilities of what we’ve learned about LLMs,” he said.
“I think you’re going to start seeing a lot of different techniques meld together, but there’s going to be some new pieces that are going to be invented that we don’t even know yet are required to get a true AGI,” he added.
The potential of AGI development
Despite the theoretical aspect of AGI, there is an idea that the industry is preparing for the best and worst of what’s to come with the emerging technology. There’s already news of brands taking the guardrails off standard AI models as they advance in complexity. Meanwhile, there is also research evidence revealing that AI models can be intentionally deceptive to human users– a trait that is unlikely to improve as technology becomes more autonomous. However, the experts consider that humans will remain as stewards of the technology.
“The reality is I would hope there would be checks and balances in any kinds of systems. Where AI is interesting, but I wouldn’t let it run free in my company,” Olson said.
Roetzer noted that before ChatGPT the industry did not know what form AI would take. Currently, leaders are in the same space with AGI and must continue experimenting with what is available until something new is developed. The distillation method that was made notable by the Chinese AI company DeepSeek has been highlighted is the closest option to an innovation for AGI at this time.
“I think you could potentially see the same kind of thing with AGI, if history repeats itself. They’re probably going to require tons of resources specialized around that. But then we learn more about what is really needed to make it work– how it works,” Olson said.
Similar to the original Al small language model, run through a distillation process, which is trained for specific tasks on more simple GPUs or even a smartphone, a potential AGI could be trained on more modest hardware at lower production costs.
“I could potentially see a distillation style technique coming to AGI where we lose some abilities, but maybe we gain the intelligence we specifically need for that task and focus on that then people can run those locally and build those themselves as they can today with SLMs by using the vendor’s LLM where they put all the money in already now,” he added.