AI Pills: Understanding the Three Levels of AI Belief

In an era where discussions about artificial intelligence often feel fragmented and polarized, Zvi Mowshowitz offered a compelling framework on his Substack, titled 'The Three AI Pills.' This metaphor provides a simple yet potent way to categorize people's understanding of AI's true capabilities. It helps explain why, in the same conversation, one person might dismiss AI as a mere toy, while another views it as a transformative, even existential, force.

Decoding the Three Pills

Mowshowitz's 'pills' aren't products, but rather escalating levels of belief in AI's potential. Think of them as progressive stages of conviction:

• The AI Pill: This is the foundational belief — acknowledging that AI exists and can perform the tasks it demonstrably does today.
• The AGI Pill: Moving beyond current capabilities, this pill signifies a belief that AI will soon be able to do much more, approaching general intelligence akin to human cognition.
• The ASI Pill: The most advanced stage, this is the conviction that AI will, within our lifetimes, surpass human performance in virtually every domain.

Mowshowitz openly states he is 'ASI pilled,' and suggests that a significant portion of staff within leading AI labs, and the labs themselves, share this perspective. This isn't a cautious, middle-ground assertion; it's a bold bet on the most powerful end of the AI spectrum.

Why Many Haven't Even Taken the First Pill

What Mowshowitz finds most frustrating are those who haven't even swallowed the first pill. He emphasizes that such individuals are not rare; they are, in his words, 'everywhere.' Many people simply aren't aware of what cutting-edge AI can accomplish right now. They haven't experimented with sophisticated coding agents, or if they've used tools like ChatGPT, it's often for trivial tasks. Their understanding is frequently colored by memories of AI failures from years past.

These skeptics often cite outdated research, dismissing AI as a 'stochastic parrot' that merely regurgitates its training data. Mowshowowitz argues that some of these views are incorrect, while others are 'not even wrong' – meaning they lack a coherent basis for discussion. This might sound harsh, but anyone who has tried to explain current AI capabilities to such individuals, only to be met with anecdotes about a chatbot misidentifying a restaurant or suggesting the wrong pizza quantity, can likely empathize with his frustration.

The Untapped Potential of the First Pill

Even fully embracing just the first pill – acknowledging 'AI can do what it can do now' – is enough to revolutionize personal workflows. Mowshowitz points out that many tasks previously done manually can now be accomplished with a quick text prompt. Things that weren't worth doing before suddenly become feasible. Questions that were too much effort to ask can now be posed effortlessly. The marginal cost of trying AI to assist with tasks is often near zero.

He also notes that most economists, as well as those in policy and government, typically only reach the first pill, often underestimating its profound impact. For instance, some might dismiss AI for a specific application as 'too expensive,' failing to grasp that equivalent capabilities will soon become orders of magnitude cheaper. This observation highlights a critical disconnect between those building and deploying AI and those shaping its regulatory and economic landscape.

The true value of Mowshowitz's commentary isn't in offering specific tool recommendations or technical predictions. Instead, it provides a crucial coordinate system for understanding where someone stands in the AI debate. The next time you find yourself arguing whether AI is a bubble or a revolution, you might first ask: 'How many pills have you taken?'

Of course, this framework has its limitations. It compresses complex positions into three distinct categories, overlooking nuances like specific timelines for capability development or differing directions of AI progress. However, as an introductory framework for dissecting AI disagreements, it's remarkably sharp.

If you frequently engage in AI-related discussions, these three pills can serve as a quick classification tool. By first confirming which level of reality your interlocutor acknowledges, you can better decide whether to delve deeper. Many unproductive arguments stem from two people, on different 'dosages,' talking past each other.