Source URL: https://www.theregister.com/2025/03/05/boffins_from_china_calculate_agi/
Source: The Register
Title: Worry not. China’s on the line saying AGI still a long way off
Feedly Summary: Instead of Turing Test, subject models to this Survival Game to assess intelligence, scientist tells The Reg
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.…
AI Summary and Description: Yes
Summary: The text discusses a novel approach by researchers from Tsinghua University and Renmin University, who have proposed a method called the Survival Game to determine if AI models can qualify as Artificial General Intelligence (AGI). This approach evaluates whether AI can autonomously solve problems through trial and error, a vital capability for real-world applications. The significant challenges related to achieving AGI, including prohibitive hardware and computational costs, are also highlighted.
Detailed Description:
– **Introduction to AGI and the Turing Test**: The discussion begins with a historical context, referencing Alan Turing’s Imitation Game that evaluates AI’s ability to mimic human responses convincingly. Currently, the industry is exploring AGI, a theoretical stage of AI where machines can understand and learn tasks similarly to humans.
– **The Survival Game**:
– Researchers in China have developed a method called the Survival Game to test AI systems for signs of AGI.
– This test leverages concepts from natural selection, determining if AI can ‘survive’ by solving problems through trial and error.
– If an AI system solves a problem within a limited number of attempts, it ‘survives’ and can progress; if it fails, it must be retrained.
– **Evaluation Domains**: The Survival Game evaluates AI across various domains:
– **Image Classification**: Measures the trial-and-error attempts for correct classifications.
– **Question Answering**: Tests models against datasets like MMLU-Pro, NQ, and TriviaQA.
– **Mathematical Problem Solving**: Assesses performance using datasets such as CMath and GSM8K.
– **Resource and Cost Considerations**:
– The researchers point out substantial obstacles in achieving AGI, particularly concerning hardware limitations and costs:
– They project needing 1026 parameters to pass AGI tests, vastly exceeding the human brain’s neuron count.
– Supplying the necessary computational resources (5 × 10^15 GPUs) would be financially implausible, costing far beyond the market value of major corporations like Apple.
– **Challenges and Implications**:
– Despite advancements, current AI struggles with problems involving trial-and-error learning, impacting real-world applications in areas like autonomous driving.
– The researchers argue for the necessity of breakthroughs in AI technology and hardware to reach a feasible AGI state.
– **Conclusion**:
– This study suggests both optimism and caution in the journey toward AGI, underscoring the significance of adaptability in AI systems to confront real-world complexities effectively.
– The discussion encourages critical evaluation of methodologies in measuring AI progress, acknowledging both advancements and systemic limitations.
The insights from this research carry profound implications for security and compliance professionals, particularly in developing robust frameworks to evaluate AI systems and ensuring they meet both ethical standards and operational effectiveness in diverse applications.