The most important century (with Holden Karnofsky)

Aug 27, 2025 Episode Page ↗
Overview

Holden Karnofsky, a Member of Technical Staff at Anthropic and co-founder of GiveWell and Open Philanthropy, discusses whether society is at "peak progress," the dynamics of innovation, and the radical uncertainty of an AI-driven future. He explores the implications of exponential growth, the "low-hanging fruit" theory, and the critical need for careful AI development and pandemic preparedness.

At a Glance
20 Insights
1h 48m Duration
10 Topics
5 Concepts

Deep Dive Analysis

The Special Nature of Our Current Era

Limits to Exponential Growth and Innovation Stagnation

Evaluating Historical and Present Intellectual Productivity

Technology's Mixed Impact on Quality of Life

Risks from Advanced Technologies and Unlearned Lessons

Navigating AI Risks: Policy, Technical, and Personal Shifts

AI Alignment, Emergent Goals, and Rapid Advancement

Holden's Approach to Learning and Decision-Making

Effective Altruism: Values, Impact, and Neglected Areas

Update on Psychology's Reproducibility Crisis

Special Time in History

The current era is unique due to unprecedented global economic growth and technological acceleration over the last 200-300 years, making it difficult to predict the future based on past trends. This period contrasts sharply with previous millennia of very low growth.

Low-Hanging Fruit Dynamics

This principle suggests that innovation becomes harder and less productive per researcher over time in any given field because the easiest and most foundational discoveries are made early on. This explains why early innovators often have outsized impact and why per-head output decreases as a field matures.

Success Without Dignity

This concept describes a scenario where humanity achieves a good outcome with advanced AI despite a 'pathetic response' or lack of responsible, coordinated effort from society. This could involve AIs being 'human-like' in their values and capabilities, or rapid technical progress in safety measures during a short window after AGI development.

Learning by Writing

This is a personal learning methodology where one starts by articulating their current opinion on a topic, then uses that as a basis to identify research questions, read critically, and seek external critiques to refine and update their understanding. This is contrasted with extensive reading before forming an opinion.

Evolution as Misaligned AI

This analogy suggests that natural selection, like a programmer, 'trained' organisms (humans) to maximize offspring. However, humans developed emergent goals (happiness, avoiding children) that are not perfectly aligned with the 'programmer's' original objective, illustrating how AI could develop unintended goals.

?
Why do we live in a special time in history?

The last 200 years since the Industrial Revolution have seen unprecedented global economic growth and technological change, clumping significant events and accelerating growth rates far beyond any previous period in human history.

?
Can exponential growth continue for thousands of years?

No, a few percent annual growth for 10,000 years would require cramming hundreds of times the world's current economic value into every atom in our galaxy, which is physically impossible due to constraints like the speed of light.

?
Why does innovation per researcher seem to decline over time?

This is explained by 'low-hanging fruit' dynamics; early in a field, the easiest and most foundational discoveries are made, making subsequent breakthroughs harder and requiring more resources per unit of output.

?
Did quality of life improve with the agricultural revolution?

Evidence suggests quality of life may have worsened, with increased violence (homicides per capita) and potentially poorer nutrition (indicated by height) in sedentary agricultural societies compared to roving hunter-gatherer bands.

?
How does Holden Karnofsky approach learning new topics?

He uses a 'learning by writing' method, starting by articulating his current opinion, then identifying research questions, reading critically to challenge his views, and seeking external critiques to refine his understanding.

?
What is Anthropic's approach to managing AI risks?

Anthropic focuses on designing a responsible scaling policy, securing systems for more powerful AI, and planning to reduce the risk of extreme human abuse of AI, in addition to the risk of AI taking over for itself.

?
What is the 'human-level AGI' threshold, and why is it significant?

This threshold is when AI can perform the kind of work done by the best human AI research teams. It's significant because once AI can do AI research on its own, it could lead to extremely rapid, self-accelerating improvements in AI capabilities.

?
How does the concept of 'misaligned AI' relate to human evolution?

Humans can be seen as 'misaligned AI' from natural selection's perspective; while evolution 'programmed' us to maximize offspring, we developed emergent goals (like happiness or avoiding children) that diverge from this original objective, illustrating how AI could develop unintended goals.

?
What is Holden Karnofsky's view on metaethics and philosophy?

He is skeptical of philosophy's methodology and its ability to provide definitive answers on what we should value. He gives little weight to philosophical thought experiments when they conflict with his intuitions, heuristics, or 'juju.'

?
Are there still neglected areas for doing good, despite increased attention from effective altruism?

Yes, Holden believes areas like AI safety, pandemic preparedness/biosecurity, animal suffering on factory farms, and global poverty are still massively neglected and offer incredible opportunities for impact.

?
What is the current state of psychology paper reproducibility?

While about 40% of top psychology papers 15 years ago might not replicate, current replication efforts show a much lower failure rate (e.g., 2 out of 12 for Clearer Thinking's project). However, new issues like 'importance hacking' and incorrect statistics are still observed.

1. Embrace Epistemic Humility

Practice admitting when you are wrong to build resilience and improve long-term well-being and epistemic accuracy. Cultivate a preference for discovering and correcting errors, even if initially embarrassing, over maintaining false beliefs.

2. Learn by Writing

Adopt a “learning by writing” approach by articulating your current opinion first, then using it to identify research questions and guide your learning. Externalize and rigorously test your ideas by writing them down, outlining evidence, and actively seeking counterarguments before fully committing to a stance.

3. Read Non-Fiction Strategically

Read non-fiction strategically by skimming, focusing on introductions, and engaging with criticisms to guide deeper dives into relevant sections. This approach helps retain more information relevant to your specific interests.

4. Form Flexible Opinions

Form opinions early, even with limited knowledge, but hold them with varying strengths and be open to easy revision. Distinguish between forming an opinion (acceptable with limited knowledge) and acting forcefully or making high-stakes decisions based on it (requires more expertise).

5. Vet Expert Knowledge

Develop a strategy for identifying trustworthy experts by doing foundational research yourself, then deferring to them on specific topics. Adjust your learning investment based on the field’s complexity and your need for informed judgment.

6. Acknowledge Future Uncertainty

Acknowledge the radical uncertainty of the future by understanding historical growth patterns, and avoid extrapolating future trends solely from recent history. Recognize the current era as historically unique due to rapid technological and quality-of-life changes.

7. Plan for Stagnation

Prepare for potential stagnation or collapse rather than assuming continuous exponential growth, as indefinite multi-percent growth is unsustainable. Understand that innovation per researcher tends to decrease over time due to “low-hanging fruit” being picked.

8. Accelerate Innovation via Minds

To accelerate innovation, significantly increase the number of minds working on problems, potentially through AI, to outweigh low-hanging fruit dynamics.

9. Nuance Progress’s Impact

Maintain a nuanced view on progress, acknowledging that while growth has generally been good, new technologies like AGI could have unforeseen negative consequences. Reframe the perception of historical “geniuses” by understanding that modern intellectual talent is abundant but faces different innovation landscapes.

10. Mitigate New Addictions

Be aware that technological progress can create new forms of addiction and “traps” that exploit human psychology, such as social media and processed foods.

11. Societal Problem Mitigation

Recognize that societal problems caused by technology can often be mitigated through collective action and regulation, as seen with air pollution.

12. Focus AI/Bio Safety

Focus safety efforts on specific, foreseeably high-risk technologies like bioweapons and AI, rather than broadly restricting all innovation. Pursue both technical solutions and regulatory frameworks for AI safety, as they are complementary and mutually reinforcing.

13. Learn from Pandemics

Do not assume society will automatically learn from and effectively respond to major crises like pandemics without deliberate effort. Advocate for low-cost, high-impact interventions (e.g., improved air circulation) in public health responses.

14. Identify AI Research Threshold

Identify AI’s ability to autonomously conduct high-level AI research as a critical threshold for exponential progress and heightened risk, signaling a need for extreme caution.

15. Separate Responsibility, Outcome

Differentiate between responsible action and achieving a good outcome, recognizing that positive results can sometimes occur despite irresponsible approaches.

16. Prioritize Career Fulfillment

When giving career advice, prioritize what individuals genuinely have energy for and can excel at, rather than solely focusing on external impact metrics. Approach complex, high-stakes work like AI development with positive energy and high integrity, as personal well-being correlates with a reduced risk of inadvertently causing harm.

17. Strategic Impact Allocation

When evaluating impact, prioritize direct, comparable benefits (apples-to-apples) and seek to maximize them. Balance quantitative comparisons with intuitive diversification when comparing incommensurable benefits to avoid excessive noise and negative consequences.

18. Skepticism of Pure Philosophy

Use philosophical thought experiments to challenge intuitions, but prioritize your core heuristics and values over purely theoretical philosophical arguments in high-stakes decisions. The methodology of philosophy is often not robust enough to outweigh practical judgment.

19. Support Neglected High-Impacts

Direct attention and effort towards highly neglected but impactful areas such as AI safety, pandemic preparedness, animal welfare, and global poverty. These fields offer immense opportunities for doing good that are currently under-resourced.

20. Value Modern Social Science

Prioritize learning from contemporary social science for higher quality reasoning and understanding, noting its consistent improvement. Modern social science, particularly in causal inference, has become significantly more rigorous and reliable than in past decades.

If you look at the whole trend and extrapolate it, you would project explosive growth. And there's been papers making that argument. You could also expect things to collapse.

Holden Karnofsky

I think when people imagine what the future is going to be like, they think about all they've ever known. But all they've ever known is a tiny fraction of human history.

Holden Karnofsky

It's like, I don't know, like humans are pretty different from each other. We're pretty different from ourselves or we're confusing or we're inconsistent. There's a lot of stuff we want. There doesn't seem to be any one thing that we strategically prioritize over everything else that varies by the human. It's a little hard to predict.

Holden Karnofsky

I feel like natural selection is a bit of a worst case for how you would train an AI.

Holden Karnofsky

I think philosophy is an extremely unimpressive field, in my opinion, the methodology just isn't very good. And there's not much reason to think we've like learned a lot from doing philosophy.

Holden Karnofsky
3 million years or 300,000 years
Humanity's age (depending on account) Depending on the specific account of human history
10,000 years
Human civilization's age Approximate age of human civilization
Almost never 1% per year
Global economic growth rate before Industrial Revolution Prior to the last 200-300 years
A few percent per year
Global economic growth rate since Industrial Revolution Consistent growth rate in recent centuries
Factor of 2-3
Variation in output per head in wealthy, educated countries Typical range of variation, not a factor of 10
6 to 12 months
Time from human-level AGI to superintelligence (Holden's guess) A very long time in terms of research potential
About 40%
Psychology papers failing to replicate (approx. 15 years ago) In top psychology journals, if good replication studies were attempted
12
Clearer Thinking's psychology replication attempts Only 2 out of 12 failed to replicate, with different reasons than expected