The world is awakening to the reality that intelligence will be automated sooner than we ever imagined. However, we’re still grappling with who will reap the rewards. This article explores the latest developments in Artificial General Intelligence (AGI), drawing insights from Sam Altman’s recent essay and other industry perspectives.
The Shifting Sands of AGI Definition
Altman offers yet another definition of AGI: a system capable of tackling increasingly complex problems at a human level across many fields. And, by this definition, we’re rapidly approaching that reality.
AGI is surpassing Human Capabilities.
Consider coding: OpenAI’s O3 model was once ranked among the top 175 coders in Codeforces ELO. Fast forward, and they now have an internal model scoring within the top 50. These systems aren’t merely mimicking top performers; they’re experimenting and learning through reinforcement learning. This progress extends beyond coding. OpenAI’s Deep Research tool has even assisted in suggesting diagnoses for medical cases, identifying insights that human doctors might miss. While it’s not perfect and may hallucinate, the potential is undeniable, especially as we move towards systems like O5 searching exponentially more sources.
Karina Nenn from OpenAI speaks on task saturation. Reinforcement learning allows models to learn infinite tasks, such as searching the web, using a computer, or writing well. According to her, the bottleneck is now in evaluations. This is seen with tasks like online purchases or spreadsheet completion, which are verifiable and easily optimized through reinforcement learning.
The Exponential Investment in AGI
Altman emphasizes that the intelligence of an AI model is roughly proportional to the logarithm of the resources used to train and run it. The socioeconomic value of each incremental increase in intelligence grows super-exponentially. Therefore, there is no reason for the exponentially increasing investment to stop.
Altman has previously suggested that OpenAI could capture a significant portion of the world’s wealth through AGI creation and redistribute it. He speculated figures reaching trillions or even hundreds of trillions. This scale could potentially impact the entire global labor force.
The Battle for AGI Control
Elon Musk has bid almost $100 billion for OpenAI’s nonprofit arm, challenging Altman and Microsoft’s control. The nonprofit’s stake in OpenAI is valued at around $40 billion by Altman and his team, leading to potential court battles and equity dilution.
The Promise of Cheaper Goods (and More Expensive Luxuries)
With AGI, Altman predicts a dramatic drop in the price of many goods. However, luxury goods and land may become even more expensive. He also hinted at an upcoming hardware device designed with Johnny Ive from Apple.
In contrast, smaller language models might become more accessible. Altman mentioned the possibility of open-sourcing “mini” versions of future models like O4.
AGI for All of Humanity?
OpenAI’s mission is to ensure AGI benefits all of humanity. However, achieving this goal is challenging, especially with the potential for widespread job displacement. Yoshua Bengio, one of the leading pioneers of AI, raises concerns about nations or companies using AGI to undermine other economies. He suggests that powerful entities might hoard the most advanced AI systems, widening the gap between those who have access and those who do not.
The Rise of Competitive AI Models
While models like Gemini 2 Pro and Flash from Google Mind show promise, their benchmark results are not yet at the level of O3 or DeepSeek R1. However, Gemini excels at quickly reading vast amounts of PDFs and other files at a low cost. As ChatGPT continues to gain popularity, Google is expected to invest heavily in improving Gemini’s capabilities.
The Dark Side of AGI: Control and Surveillance
Altman warns of the potential for authoritarian governments to use AI for mass surveillance and control. A Rand Corporation paper highlights threats to national security, including:
- Wonder Weapons
- Systemic shifts in power
- Non-experts wielding weapons of mass destruction
- Artificial entities with agency
- Instability
The paper also suggests that the U.S. is not well-positioned to realize the economic benefits of AGI without widespread unemployment and societal unrest.
AGI for the Masses?
Stanford researchers have demonstrated that even small, open-source models can achieve impressive results with limited data and compute time. By manipulating token weights, they were able to train a 32-billion parameter model to compete with O1 on challenging reasoning tasks. This suggests that AGI capabilities may become more accessible to individuals and smaller organizations.
Conclusion
The path to AGI is fraught with both immense promise and potential peril. As we move closer to this reality, it’s crucial to address the ethical, societal, and economic implications. Ensuring that AGI benefits all of humanity will require careful planning, collaboration, and a commitment to equitable distribution of its rewards.