Worry not. China's on the line saying AGI still a long way off

In 1950, Alan Turing proposed the Imitation Game, better known as the Turing Test, to identify when a computer's response to questions becomes convincing enough that the interrogator believes the machine could be human.

Generative AI models have passed the Turing Test and now the tech industry is focused on Artificial General Intelligence (AGI), the hypothetical point at which a computer can understand or learn any intellectual task as well as a human.

Presently, AGI is vaguely defined and does not exist, though there are people already trying to prevent its emergence. Among AI boosters, AGI is a bit like quantum computing – a distant goal cited for funding.

Citing intelligence tests devised by Turing and others – though disappointingly not the Voight-Kampff test from Blade Runner – researchers in China have proposed a method called the Survival Game to determine whether AI models qualify as AGI.

The Survival Game is essentially a simplified form of natural selection

Authors Jingtao Zhan, Jiahao Zhao, Jiayu Li, Yiqun Liu, Bo Zhang, Qingyao Ai, Jiaxin Mao, Hongning Wang, Min Zhang, and Shaoping Ma – affiliated with Tsinghua University and Renmin University of China – describe their approach in a preprint paper titled: "Evaluating Intelligence via Trial and Error."

"The main idea behind this paper is to assess whether current AI systems can find solutions through continuous trial and error," Jingtao Zhan, a PhD student in computer science at Tsinghua University and corresponding author, told The Register.

"If an AI system can find a solution within a limited number of attempts, it is considered to 'survive'; otherwise, it 'goes extinct.'"

Models that survive are allowed to progress to other tests; ones that don't pass get retrained until they do, which is a significant process.

The Survival Game covers various knowledge domains. In image classification, for example, the test assesses how many trial-and-error attempts are required before the model comes up with a correct classification. In question answering, models are tested against three well-known datasets: MMLU-Pro, NQ, and TriviaQA. In mathematics, the test measures performance using three math datasets: CMath, GSM8K, and the MATH competition dataset.

Supporting code has been published to GitHub.

"The Survival Game is essentially a simplified form of natural selection, and we aim to use this approach to test whether AI can adapt and learn through such a mechanism," said Zhan.

"If an AI system passes this test, it means it can autonomously find solutions without human supervision and operate independently. This serves as both my perspective on AGI and a way to evaluate it."

• Why making pretend people with AGI is a waste of energy
• As China embraces Big Tech again, Alibaba plans vast spend to push for artificial general intelligence
• Microsoft warns Trump: Where the US won't sell AI tech, China will
• Phantom of the Opera: AI agent now lurks within browser, for the lazy

The researchers' results suggest that even if Moore's Law – the projected doubling of chip transistor density every two years – were to continue beyond its arguable demise in 2016, the cost to build a neural network capable of passing the above AGI tests would be exorbitant and it would take 70 years for hardware to be able to support the anticipated model.

"Projections suggest that achieving the autonomous level for general tasks would require 10²⁶ parameters," the paper says.

That's a huge number: "Five orders of magnitude higher than the total number of neurons in all of humanity’s brains combined," the authors observe, where a human brain has 10¹¹ neurons and population is approaching about 10^10 people for a neuron total of 10^21.

Setting aside computation costs such as training and inference, just loading a model with that many parameters onto Nvidia H100 GPUs would be an untenable extravagance.

They struggle significantly when faced with problems that require continuous trial and error to find solutions

"Since the memory of an H100 GPU is 80GB, we would need 5 × 10¹⁵ GPUs," the paper says. "Based on the cost of H100 GPUs ($30,000) and the market value of Apple Inc ($3.7 trillion) in February 2025, the total value of these GPUs would be equivalent to 4 × 10⁷ times the market value of Apple. As we can see, without breakthroughs in hardware and AI technology, it is infeasible to afford scaling for autonomous-level intelligence."

Zhan argues these results indicate AI technology has a long way to go before it can autonomously solve unknown problems, particularly in an open environment where it must adapt through natural selection.

"While current AI systems may perform well in certain benchmarks, achieving high accuracy in predefined tasks, they struggle significantly when faced with problems that require continuous trial and error to find solutions," said Zhan.

The study, Zhan observes, shows that when AI models fail, they rarely adapt to come up with a correct solution through iterative attempts.

"In the Survival Game, this means it cannot survive," said Zhan. "Such trial-and-error learning is crucial in real-world applications, particularly in areas like tool use, autonomous agents, and self-driving cars. If AI can truly learn to solve problems through trial and error, it will mark a significant step toward widespread real-world deployment."

Food for thought. Whether you agree with the team's methodology and approach or not, and some of us here are a little skeptical of the study, we welcome people trying to calculate the trajectory of AI technology without the hype or grift. ®