AGI Definition Threatens Microsoft-OpenAI's Multi-Billion Deal

A multi-billion dollar partnership between Microsoft and OpenAI is reportedly undergoing significant renegotiation, driven by the elusive definition of artificial general intelligence (AGI). The existing agreement, valued at a reported $13.75 billion, grants Microsoft rights to OpenAI's models until 2030, or until OpenAI declares it has achieved AGI. This clause, once seemingly theoretical, has become a critical legal boundary as AGI transitions from science fiction to a tangible possibility.

Should OpenAI declare the achievement of AGI under current terms, Microsoft risks losing access to the foundational technology upon which much of its future strategy rests. This includes key products and initiatives such as its Azure cloud platform, Office software suite, GitHub development tools, and the generative AI ambitions of Copilot. The immense stakes have spurred intensive discussions between the two AI powerhouses.

According to reports from Bloomberg and TechCrunch, Microsoft is seeking to secure continued access to OpenAI's technology even after AGI is achieved, potentially by acquiring a larger equity stake in OpenAI, rumored to be in the low to mid-30% range. OpenAI, for its part, is pushing for greater autonomy. This includes fewer restrictions on how and where it can sell its models, stricter controls over Microsoft’s deployment of its technology, and a larger share of the revenue.

Adding to the complexity, OpenAI is also navigating its transition towards a fully commercial structure. Currently operating as a nonprofit that governs a capped-profit company, a model designed to limit investor returns, this structure is increasingly seen as unsustainable amid swirling IPO rumors and massive infrastructure demands.

The negotiations have not been without friction. Microsoft has reportedly blocked some of OpenAI's attempted acquisitions, signaling its own expanding AI ambitions. With Mustafa Suleyman, co-founder of DeepMind and Inflection AI, now leading Microsoft AI, the company appears to be positioning itself to potentially hedge against, or even directly compete with, OpenAI. Further complicating matters, OpenAI is reportedly seeking additional compute resources beyond what Microsoft can provide, exploring potential deals with other providers like Google and Oracle, indicating a desire to avoid being confined.

At the heart of these high-stakes discussions is the lack of a clear, universally accepted definition for AGI. For the purpose of their agreement, Microsoft and OpenAI must arrive at a mutually agreeable standard. One benchmark frequently cited in industry discussions defines AGI as an AI capable of outperforming humans in most economically valuable work.

However, the practical application of such a definition remains a challenge. An emerging concept gaining traction in the broader AI industry is the Economic Turing Test, proposed by AI safety researcher and Anthropic co-founder Ben Mann. As articulated on Lenny’s Podcast, this test suggests that if an AI agent can perform a job for one to three months and is hired over a human without the employer realizing it’s a machine, it passes. Extending this, Mann suggests that if AI passes this test for 50% of "money-weighted" jobs, it would signify the advent of transformative AI, potentially leading to a substantial increase in GDP growth. Mann has projected this could occur as early as 2027 or 2028.

While the definition of AGI is still being debated in boardrooms and research labs, some observers suggest that AGI-level performance may already be subtly influencing the economy. Instances where individuals leverage AI to complete job assessments, cheat hiring workflows, or offload actual work to AI agents behind the scenes are cited as examples where employers might unknowingly be hiring or utilizing AI, a strange twist on the Economic Turing Test.

These complex negotiations underscore the profound implications of AGI's potential emergence, not just for the future of two of the world's most influential tech companies, but for the broader economy and the very definition of human work.