Million Dollar Questions

What if Warren Buffett donated 50% of his wealth to AI compute?

OpenAI launched their new o3 model last week. It barely made the WSJ frontpage, as if it was just another day in Silicon Valley. But buried in that technical announcement was something extraordinary: the first glimpses of how AGI might actually emerge.

(credit @ben_j_todd)

Early benchmarks show that o3 represents a significant leap forward. But this isn't just another "smarter" language model. The fundamental shift lies in how it approaches problem-solving through massive parallelization and search over the space of possible solutions.

The o3 model combines two key elements. First, there's the foundation model capable of generating thousands to millions of potential solutions for any given problem. These "chains of thought" represent different paths to reaching an answer. Second, there's an evaluation system that determines which of these chains is optimal. While this principle isn't new — DeepMind's AlphaZero used a similar approach for games like chess and go — o3 elevates it to a higher level of abstraction. It's not just about solving specific problems anymore; it's about automating the entire thinking process.

To better understand this: The original GPT-4 is a smart model that provides a single answer to each chat query. Its successor, o1 (released three months ago), features a slightly enhanced foundation model but primarily utilizes serial processing called a chain of thought. The model first plans its solution steps, then executes them. That's why each response takes anywhere from tens of seconds to several minutes. Now, o3 launches dozens, hundreds, or thousands of such chains of thought simultaneously (depending on what you're willing to pay) and selects the best one. In essence, it's exploring the space of potential thoughts – the landscape of possible solution paths.

This has several significant implications. Clearly, the era of cheap artificial intelligence is over, at least temporarily. Until recently, we paid $20 monthly for access to cutting-edge models, but the future points toward substantially higher costs. OpenAI already offers a Pro version at $200 monthly, and o3 might cost hundreds or thousands of dollars per query¹, as massive parallel processing demands enormous computational resources.

The proverbial "million dollar question" thus takes on literal meaning. Which answer would you pay a thousand dollars for? Which answer would humanity pay a million dollars for? Or a hundred million? At every level of society – individuals, companies, organizations, nations – we immediately encounter resource constraints.

The intelligence discussion is transforming into a question of efficiently utilizing available computational capacity. It will be fascinating to see which questions take precedence. How will science change? There are already reports of o1 accomplishing in minutes what took a PhD student months – and o3 will be even more powerful. Research directions will increasingly be determined not by grants but by which questions people or organizations are willing to pay for. How will philanthropy evolve? Will foundations shift from allocating money to nonprofits toward funding complex computational queries and answers? What could be solved next year, if Warren Buffett donated 50% of his wealth to AI compute? While today's charitable campaigns raise money for expensive treatments for rare diseases, soon they might collectively fund AI compute to discover entirely new medicines.

Intelligence will function less like a muse and more like a market.

The o3 model thus clarifies our path to AGI. We won't see a sudden arrival of cheap, omnipotent superintelligence – instead, we'll likely see a gradual process constrained by computational costs, much as physical laws constrain the material world.

Perhaps a new discipline will emerge, a kind of 'intelligence economics,' addressing the crucial question: Which problems are important enough to warrant solving at extreme costs of thousands or millions of dollars per query? It's not just about who can afford to pay the most. The strategic questions will be key: which answers will help reduce future computation costs? Which discoveries will unlock paths to solving other problems?

The current situation recalls the Manhattan Project, but in a more positive sense. It wasn't just about having brilliant physicists on the team. What won the war was their strategy. They had to systematically evaluate which research paths were most promising because they lacked the time and resources to explore every possibility. Today, we may face a similar challenge with AI, but with a much broader scope and more constructive goal – how and where to invest our collective intelligence.

The situation has another dimension: steep price drops. Over the coming months and years, inference costs will fall dramatically – tenfold, hundredfold. But rather than simplifying matters, this adds complexity. Which question are you willing to pay $1,000 for today, when you could get the answer for $10 in a year? What can't wait?

One such million-dollar question is clear. This new model can not only solve our problems but can likely also seek ways to improve itself. OpenAI is surely already asking o3 how to design o4. It can optimize its foundation model, evaluation system, and the entire process of finding the best thought chains. We're rapidly approaching an era where AI creates the next generation of AI. And that, so far, is the most likely path to AGI.

1. Some estimates of the cost of the o3 benchmark scores go into millions.
https://dev.to/maximsaplin/openai-o3-thinking-fast-and-slow-2g79