Is OpenAI's Astra (GPT-6) an example of Artificial General Intelligence (AGI)? Can AGI innovate business and elevate the economy to a new level? What exactly is AGI, and how does it compare to human intelligence?
Human intelligence defines and solves problems simultaneously. AGI addresses given problems independently, finding solutions without prior knowledge in unfamiliar environments, though it currently operates only in digital contexts. It lacks the understanding of the physical world, limiting its ability to effectively engage with real-world issues.
While a human can perform tasks with 20 watts of energy, current AGI systems require hundreds of megawatts and large-scale systems. Humans, constrained by energy limits, rely on experience and intuition for leaps in reasoning, whereas AGI can derive optimized answers through extensive repetition over days, producing results beyond human imagination.
Humans alone bear the responsibility for outcomes. AGI-equipped agents communicate efficiently in a language called Neuralese, encoded as vector values, while humans, despite their less efficient communication, collaborate through empathy and the creation of invisible narratives. To control AGI, it is essential to develop technologies that can decode their communications while avoiding interference.
Can we dismiss Astra as not being AGI? It has achieved a 92.7% accuracy in the ScreenSpot-Pro benchmark test, which measures the ability to accurately perceive screens in commonly used operating systems and web browsers. In the OSWorld 2.0 benchmark, which assesses the ability to complete complex tasks using a mouse and keyboard, Astra recorded a 72.6% success rate within an average of 40 minutes.
Other AI models perform at approximately 60-70% efficiency within an average of 75 minutes, while humans fall into the 72.3% range. These results reflect Astra's efficient mechanisms in problem-solving through repeated reasoning. In the GPQA Diamond benchmark for doctoral-level reasoning, Astra achieved a success rate of about 96%. Other AI models encounter hallucinations in the 70-90% range, while expert human reasoning yields an average of 65-74%.
In the ARC-AGI benchmark, which tests problem-solving in unfamiliar environments without prior knowledge, Astra reached a remarkable 99.9%. Other AI models have been stuck at 30-40% for years, while humans average around 48%.
Can AGI generate revenue? Is it possible for nations to sustain growth through AGI? Will we continue to see an era where increasing employment and system improvements lead to higher revenues? Perhaps we should ask where AGI's computational power is not needed. Is it in questioning, multi-agent management, and the ability to validate AGI outputs?
If autonomous reasoning by AGI becomes the norm, then competitiveness may hinge not on software sales or labor costs, but on the ability to solve problems effectively. This suggests a rapid shift towards an Agent-as-a-Service (AaaS) model, where clients are charged for the results achieved by agents. Will there be a future if we do not disrupt today's revenue model?
Is a competitive advantage achieved by providing proactive services to a single customer, securing closed local data, and enhancing AGI alignment? We are entering an era where challenging humanity's complex problems in real-time and boldly discarding traditional work methods is becoming common sense.
We must fully integrate AGI into manufacturing and services, merging agents previously confined to digital spaces with robotics to create Physical AI, supported by gigawatt-level infrastructure. Can we cultivate a system that produces paradigm designers who explore uncharted paths alongside AGI, rather than merely training individuals to find quick answers?
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.