Understanding AI Perspectives Through a Three-Pilled Framework

The three pills

I spend a lot of time thinking about Zvi’s three-pill framework. For me, it helps drop a lot of confusion. I’m less defensive about what someone’s saying when I understand that they’re coming from a different viewpoint. It helps with having an honest conversation about where someone is coming from, and then not talking past the other person.

The framework comes from Zvi Mowshowitz’s The Three AI Pills. Pill 1 is taking current AI capabilities seriously. Pill 2 is the AGI pill, or artificial general intelligence. Pill 3 is the ASI pill, or artificial superintelligence. These are the distinctions I’ve been using to break down articles and podcasts and understand people’s arguments.

People have different definitions for AGI. It might mean doing most cognitive tasks, most white-collar work, or being a drop-in assistant in most domains. With Pill 3, AI will at some point be smarter than humans and probably faster at most things. It doesn’t have to include recursive self-improvement, but once you take that possibility seriously, there’s a lot to follow through on. AI building AI, data centers, robotics, and robots building their own factories all have to be taken seriously.

An argument can sound Pill 3 about capabilities and then switch back to Pill 2 when talking about what happens next. Someone might say AI will pump out millions and millions of mathematical proofs, but it’s going to lack taste, or we’ll need humans to choose, or we’ll need humans to explain it. Any time someone says “we need humans to” do something, that’s an assumption I want to examine. Are humans needed because AI won’t be able to do it, or because people want it to remain human?

All of the more positive outcomes I’m talking about assume alignment, which isn’t guaranteed. Pill 3 also means much, much more risk in that world, so timelines and awareness matter. Pill 1 people are less likely to focus on those risks, while Pill 2 people will typically focus more on mundane and everyday risks. Of course, there’s a whole spectrum, but existential risk is more typical of Pill 2.5 and Pill 3. Recognizing where someone falls on that spectrum helps clarify what kind of risk they’re talking about.

Pill 1 and Pill 2

A useful Pill 1 statement is that AIs are jagged. That describes current capabilities and limitations. Ethan Mollick writes about this jagged profile alongside rapid capability progress. To me, the distinction comes when we move from describing what AI can do now to asking how long that progress will continue.

In his essay on jaggedness and bottlenecks, Mollick considers the possibility of supersmart AI that never fully overlaps with human tasks. He points to memory and learning as gaps that could be harder to solve than researchers expect. If those gaps persist, humans and AI continue to bring different abilities to the work.

That’s where I see Pill 1 tendencies in an argument that can appear more Pill 2 at face value. He recognizes capabilities up to a point, but whether those capabilities eventually cover the whole range of human work remains unresolved. Acknowledging rapid progress and taking AGI seriously are different steps. I want to know when the evidence is enough to take that next step.

Mollick’s later writing does go further. He discusses recursive self-improvement and declines to bet on a gradual, manageable transition. So this is a distinction within his arguments, and his writing has changed over time. The question I’m getting at is when uncertainty about continued progress becomes a reason to keep reasoning from the present.

How long do the remaining weaknesses look like things that are further behind, and when do they look like things these systems won’t be able to do? AI can do what it can do now, but AGI is not guaranteed. What would get someone to take AGI seriously, and what would get someone who expects AGI to reconsider?

Pill 2 and Pill 3

Pill 2 can already be a lot more radical than people realize. Huge populations of AGI workers, automated AI research, huge amounts of compute being built, and concentration of power can already describe a very different world. There’s still another question about whether that produces systems far beyond human intelligence.

The recent conversation between Dylan Patel and Dwarkesh helped me think through this distinction. Dylan entertains recursive self-improvement, but looks at how regulation, economic pressures, and different bottlenecks might slow the transition. Dwarkesh has more of the Pill 3 side showing, pushing through what happens when AI labour and AI research keep compounding. My reading is that Dylan puts more weight on how much the rest of the world will let that happen.

Of course there can still be bottlenecks. But I think Pill 3 finds a way, for the most part, quickly around them. That’s a judgment about how much weight to put on the bottleneck. If someone says an institution will hold everything back, I want to know what that institution looks like after AI has become much better at planning, coordination, persuasion, and finding other ways to get things done.

Sam Altman is already talking about Pill 3 capabilities. His Gentle Singularity essay includes accelerating AI research and robots building robots. What I’m talking about is his Pill 2 tone about how manageable the transition will feel. I read that as partly speaking to people where they’re at and presenting a reassuring front for OpenAI.

We haven’t yet seen the massive disruption, but there are different thresholds for usefulness, job performance, and reliability. We can keep rolling out different versions and people can keep testing them, but at some point a system can do the whole bundle of tasks. The economic effects could change very quickly at that point. I think we could go very quickly from Pill 2 to Pill 3, and that possibility needs to be carried through the discussion.

Jensen Huang’s labour forecast is very Pill 2 to me. He keeps returning to task automation and examples like radiology, where automating part of the work leaves people able to serve more patients. That can describe a stage of the transition. It still leaves the question of what happens as AI takes over more and more of the bundle.

Even within Pill 2, whole bundles of tasks can get taken over by AI. Think about psychology, which is a difficult example. There’s the clinical work, the safety, the discussion, the documentation, the following up, and the appointment setting. Different things fall at different times, but eventually covering that whole bundle is already part of what I mean by AGI.

Jevons effects depend on how demand responds when something gets cheaper. There isn’t unlimited demand for every single thing, and more demand does not by itself establish a need for more human labour. If AI can do the additional work too, the argument about employment needs another step.

Bill Gates goes further toward Pill 3 in his discussion of capabilities. He expects AI to substitute for human cognition, do physical work, and eventually function without a human checking in on it. But his proposal to preserve certain jobs for humans, including human leadership in education and mental health care, raises another question for me. Who gets to decide that those roles should remain human?

Some people will want things to be human just for the sake of being human. Others may want AI to do those things because they expect better outcomes. Preserving the option of human involvement and requiring human involvement lead to different ways of living. I think people should be able to choose different degrees of how much AIs do for them.

Francis Su’s The Enduring Value of Math, in an Age of AI helped me think through a related question. He discusses the values people develop through learning mathematics, including persistence, creativity, and being comfortable while stuck. I see the point, and I got some of that out of graduate school myself. Those values can remain even if AI does the math.

But I keep thinking about wilderness skills. They’re fun to learn, and learning them can develop values and abilities I want to have. Mathematics can have that kind of value too. That still leaves the question of how much I want to pay a prestigious institution to develop those values, especially if I can learn from a better AI teacher at some point in the future.

The same goes for the idea that people will be needed to explain AI’s outputs or decide which questions are worth pursuing. The AIs can learn how to explain their outputs, or we can train them to do that. They can learn to connect those explanations to practical situations. Why would that ability permanently belong to humans?

Diffusion and the human role

One big topic I keep coming back to is diffusion. Nathan Lambert’s recent article describes a 50-year diffusion process and expects much of everyday life to look similar even then. He also allows for compounding improvements and robotics, but expects adoption to take much longer than capability development.

Slow physical deployment is a Pill 1 observation about what AI can do now and how it’s being used. Institutions are sticky. Everyday life still looks surprisingly normal. I think that gets the present right, but I question carrying those same assumptions into a Pill 3 world.

People might say that the everyday person will never learn how to use agents. But eventually there’s an interface where you just talk, and the agents operate your phone and computer for you. They could become part of the operating system that the computer already comes with. Then the agents just work for you, and they can create the interface for you.

In that situation, adoption becomes almost automatic. People don’t have to learn a whole new way of operating the technology. AI can make the tools more usable and more applicable, and people may benefit without thinking much about AI at all.

I’m thinking about token costs going down, too. If what’s driving these gains is smarter models that are getting smaller over time, output increases while intelligence gets cheaper. Then there’s the physical side, with robots building factories and materials and production getting cheaper. That may come years later, but the physical-world costs can change too.

Damon Binder’s industrial-growth work gives a concrete version of this. He examines how AGI could operate machines and help automate physical production, with the resulting output reinvested into more production. He explicitly sets aside recursive self-improvement to superintelligence in part of this analysis. Even AGI alone could make the physical economy change much faster than people expect.

Midjourney’s imaging project is an example of the kind of change I’m thinking about. The company is developing an ultrasound scanner and plans to open a consumer spa in 2027. Its initial product is intended for body-composition imaging, with diagnostic capabilities requiring further development and review. The project is still being developed, but it already puts imaging in a different setting from the usual hospital pathway.

Now imagine that imaging eventually gets cheap enough that you can step inside for $20 and get clinically useful information. Suppose the alternative is paying $2,000 for a hospital scan, with the appointment, interpretation, and other costs built into that pathway. If a future AI can reliably interpret the cheaper scan and help send the findings to my doctor, why would I need the whole insurance and hospital system for that particular step?

Those prices are my hypothetical, but there’s already so much in the institutional pathway that could be cheaper. Then, as intelligence gets cheaper and helps improve the hardware and production, more of the physical costs can come down. If something becomes reliable enough that people are willing to try it, and they can get it and access it, some people will go around the institutions and use it.

That’s why so much of the diffusion discussion feels like a Pill 2 argument to me. It accepts capabilities going up while carrying forward assumptions about the speed of adoption and change. Is diffusion itself something that can be overcome from a Pill 3, ASI standpoint? I think removing a lot of those bottlenecks is part of the picture.

The second main point is keeping humans in the loop. When AI is getting faster and smarter and better, will the default become humans out of the loop? We’ll gradually be disempowered, but some of that could involve personally giving up decision-making because AI is really smart, and deciding how, when, and if we want to do that.

If I have a life-threatening illness and it’s difficult to understand the risks, treatments, and options, I can imagine wanting to lean heavily on AI. Suppose a future system has far more relevant knowledge than any individual doctor and can give me a reliable account of the options and their chances of success. I’m going to want to follow advice that helps treat the disease and gives me more years of life. Choosing that could be a reasonable decision for my well-being.

But then we have to take the AI’s word for it. We have to allow it to explain things to us, and we have to trust that it’s explained them the right way. That problem also comes up in alignment, including proposals to use less capable AIs to help supervise more capable AIs. Having an AI explain another AI’s decisions still leaves our understanding dependent on AI.

With enhancement, we want the ability to work alongside AIs and independently check what they’re doing. Anything that increases our understanding would help. But unless we’re willing to accept the AI’s ability to translate things down to us, we have to be able to meet that intelligence. Whether that means AI literally inside humans, I don’t know. To me, that’s what truly staying in the loop would require.

I mean both understanding and authority. People should be able to decide how much they want decisions about their lives passed on to AI, and preserve ways of living that include different degrees of that. But having the final say becomes a different thing when we can’t independently evaluate the decision. I’m going to write a whole article about this, because delegation, disempowerment, and enhancement each bring their own questions.

Examining arguments using a pilled perspective means recognizing the underlying arguments and what logically fits within each pilled layer. There isn’t anything inherently wrong with those perspectives. They just have different ways of basing knowledge, predictions, and worldviews. Of course, I think Pill 3 is the most acceptable viewpoint based on progress and what logically follows even from the currently anticipated compute buildout.

I hope this approach helps you recognize where people are coming from, make sense of their arguments, and have conversations without talking past each other.

Process note: AI was used as a co-writing, research, editorial, and publishing collaborator. Approximately 95% of the ideas are mine, while ChatGPT supplied or substantially rewrote approximately 30–40% of the draft’s wording. These are rough editorial estimates, not measured text comparisons.