Pathways to AGI
This paper critically examines the socio-political and economic contingencies behind current AI development by analyzing five key questions regarding historical pathways, decision nodes, and model trajectories to propose transparent and sustainable alternatives to the concept of Artificial General Intelligence.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Big Picture: It's Not Just About "Smart" Robots
Imagine you are watching a race to see who can build the most intelligent robot in the world. Most people think the winner will be the one with the biggest brain or the fastest processor.
This paper argues that the race isn't actually about who has the "smartest" robot. Instead, it's about who built the track, who owns the starting line, and who decided the rules of the race.
The authors say that the current "Artificial General Intelligence" (AGI) we see today isn't a natural scientific discovery that just happened. It is a specific product shaped by money, business deals, legal contracts, and the personal histories of a small group of people. They call this a "Critical Pathway Analysis"—basically, looking at the history of how we got here to see which doors were opened and which were slammed shut.
1. What is AGI? (The Moving Goalposts)
The paper starts by saying there is no single definition for AGI. It's like trying to define "a good car" when one person says it means "fast," another says "safe," and a third says "cheap."
- The Vendor Definition: Big tech companies (like OpenAI, Google, Microsoft) define AGI based on what makes them money. For example, OpenAI says AGI is when AI can do "economically valuable work" better than humans. It's like a car company saying, "We have a perfect car when it can deliver pizza faster than a human."
- The Money Definition: There are leaked documents suggesting that for Microsoft and OpenAI, AGI might not be declared until the technology generates $100 billion in profit. This turns AGI into a financial target rather than a scientific milestone.
- The Human Definition: The paper argues that humans aren't just "smart" in one way. We have different types of intelligence (like being good at math, good at understanding feelings, good at fixing things). Current AI is very good at one thing (talking and writing) but terrible at others. The authors suggest we should look for "Artificial Multiple Intelligences" (AMI)—a system that can handle many different types of problems, not just chat.
2. The "Viable System" Analogy
The authors use a theory called the Viable System Model (from a man named Stafford Beer) to explain why current AI isn't truly "general" yet.
- The Analogy: Imagine a human body.
- The Hands (System 1): These are the current AI models. They are incredibly fast at typing, coding, and summarizing text. They are the "hands" doing the work.
- The Brain & Nervous System (Systems 2-5): These parts manage coordination, safety, long-term planning, and ethics.
- The Problem: Current AI has super-fast hands, but it lacks a fully developed nervous system. It doesn't know when to stop, it can't plan for next year, and it can't understand the complex rules of society or law. It's like a body with giant, fast hands but no brain to tell it where to go or what is dangerous.
- The Conclusion: A true "General Intelligence" isn't just a smart model; it's a whole system of systems that includes the model, the rules, the safety checks, and the people managing it.
3. The Five Paths (The Story of the Race)
The paper looks at five major players and how their specific histories shaped their AI.
- OpenAI (The Corporate Turn): Started as a non-profit research lab. But they needed money, so they became a "capped-profit" company and partnered with Microsoft.
- The Pivot: They decided to keep their best models secret (behind an API) to make money, rather than sharing them with the world. This made them rich and popular, but it also meant they stopped focusing on deep safety research to keep up with product deadlines.
- Anthropic (The Safety Rebels): A group of top scientists left OpenAI because they were worried about safety. They started Anthropic to build AI that is "steerable" and safe.
- The Struggle: They tried to build a "constitution" for AI (a set of rules the AI follows). But they still needed billions of dollars from Amazon and Google to compete. The paper asks: Can you stay safe when you are desperate for investor money?
- Google (The Late Bloomer): Google invented the core technology (the Transformer) but didn't release it as a product because they were worried about safety and reputation.
- The Mistake: They waited too long. By the time they released their AI (Gemini), OpenAI had already taken over the market. Google is now trying to catch up, but their culture of "research first" clashes with the need to move fast.
- xAI / Grok (The Wildcard): Founded by Elon Musk. Their path is driven by a specific political narrative ("anti-woke") and a desire to move faster than anyone else.
- The Risk: They have a massive distribution advantage (Twitter/X) and a supercomputer built in record time. But they have almost no safety research or governance. They are the "fast car with no brakes."
- Microsoft (The Enterprise King): Microsoft didn't build the smartest AI; they built the best delivery system. They took OpenAI's tech and put it into every office computer in the world (Copilot).
- The Strategy: They focus on making AI safe enough for banks and hospitals to use. They are building the "governance" layer, even if they aren't the ones inventing the core brain.
4. The "People Factor" (The Small Circle)
The paper points out that the entire AI industry is run by a very small group of people who all know each other.
- The "Google Brain" Connection: Almost all the top AI researchers started at Google's research lab.
- The "OpenAI" Connection: Many of them left Google to join OpenAI.
- The "Anthropic" Connection: Then, a big chunk of them left OpenAI to start Anthropic.
- The Result: It's a "revolving door." The same small group of people keeps moving between companies, taking their ideas with them. This means the industry is very "insular"—they all think alike, use the same tools, and miss out on different ways of thinking (like how a biologist or a psychologist might approach the problem).
5. What Should We Do? (The Recommendations)
The paper concludes that the current path is too narrow. We are betting everything on a few companies trying to make a profit.
- Don't just watch the race: We need to fund different kinds of research, not just the "biggest model" race. We need to study how AI understands uncertainty, how it handles long-term planning, and how it fits into society.
- Make AI a Public Utility: Just like we have public water and electricity, AI should be treated as a public resource. It shouldn't just be a subscription service for the rich.
- Sovereign AI: Countries need to make sure they have their own control over AI infrastructure (computers, data, rules) so they aren't dependent on a single foreign company.
- Check the Work: We need independent auditors to check if AI is actually safe and smart, not just let the companies say, "Trust us, we are safe."
Summary
The paper argues that AGI isn't a destination we are just "discovering." It is a path we are building. Currently, we are building a path that is paved with money, controlled by a few companies, and focused on making chatbots.
To get to a truly intelligent, safe, and useful AI, we need to stop treating it like a consumer product and start treating it like a complex system that needs to be governed, audited, and built for the benefit of everyone, not just shareholders. We need to widen the road so that different types of intelligence can grow, not just the ones that make the most profit.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.