Dwarkesh Podcast: AGI is Still 30 Years Away — Ege Erdil & Tameh Besiroglu Release Date: April 17, 2025

Introduction

In this episode of the Dwarkesh Podcast, host Keshav Murugesh engages in a deep conversation with Ege Erdil and Tameh Besiroglu, founders of Mechanize, a company focused on automating all forms of work. The discussion centers around the timeline for achieving Artificial General Intelligence (AGI) and challenges the commonly held belief that AGI is imminent.

Intelligence Explosion vs. Broader Economic Transformation

Keshav Murugesh initiates the dialogue by addressing Tameh’s recent stance on the intelligence explosion, querying why the traditional concept might be misleading.

Tameh Besiroglu ([00:22]) critiques the intelligence explosion notion by likening it to an oversimplified view of the Industrial Revolution, emphasizing that technological growth is a result of multiple complementary changes across various sectors, not just a surge in "horsepower."

“It wasn't just that we had more horsepower. I mean, that was part of it. But that's not the kind of central thing to focus on when thinking about the Industrial Revolution.”
— Tameh Besiroglu ([00:50])

Timelines to AGI and Critiques of Fast Progress Perception

The conversation shifts to differing expectations regarding the arrival of AGI. While many in San Francisco anticipate rapid advancements, Ege and Tameh project a more extended timeline, estimating AGI to be around 2045 or later.

Ege Erdil ([02:05]) expresses skepticism about the fast-paced predictions based on recent AI progress, arguing that extrapolating short-term trends leads to overoptimistic estimates.

“If you look at the fraction of the economy that has actually been automated, it's very small. So if you just extrapolate that trend, you're going to say, well, it's going to take centuries or something.”
— Ege Erdil ([02:09])

Bottlenecks in AI Progress and Compute Scaling

A significant portion of the discussion delves into the technical challenges of scaling AI. The guests argue that while compute power has dramatically increased, there are diminishing returns and practical limits to scaling further.

Tameh Besiroglu ([04:05]) elaborates on the complexities of achieving AGI, noting that each new capability in AI requires substantial compute scaling and complementary innovations across different sectors.

“We might need coherence over very long horizons, or agency and autonomy, or multimodal full understanding, just like a human would.”
— Tameh Besiroglu ([04:20])

Ege Erdil ([06:40]) concurs, highlighting that the current infrastructure heavily favors AI chip production for specific applications, limiting the scalability needed for broader automation.

Historical Analogies: Industrial Revolution and Economic Growth

To contextualize AI’s potential impact, the trio draws parallels with the Industrial Revolution, emphasizing that transformative growth stems from multi-sectoral advancements rather than singular technological leaps.

Tameh Besiroglu ([12:53]) points out that while reasoning capabilities in AI seem straightforward now, developing these competencies required years of innovation and hardware improvements, akin to the gradual advancements seen during the Industrial Revolution.

The Role of Regulation and Global Coordination

A recurring theme is the importance of regulatory frameworks and how differing national policies might influence AI deployment and economic growth.

Ege Erdil ([86:36]) suggests that regulatory responses will vary globally, with some countries adopting more liberal policies towards AI, thereby accelerating growth, while others impose stricter regulations, potentially hindering progress.

“I expect heterogeneity in how different countries respond. Some will be more liberal, others less, but overall, variation will determine AI’s impact.”
— Ege Erdil ([86:36])

Tameh Besiroglu ([144:15]) emphasizes that the scale of a nation's economy and its regulatory stance will significantly influence how AI-driven growth unfolds, drawing parallels to how certain countries led the Industrial Revolution.

The Impact of AI on Economy and Society

The guests explore how AI-driven automation could lead to unprecedented economic growth by enhancing productivity across various sectors.

Ege Erdil ([172:27]) envisions a world where AI significantly boosts economic output, leading to enhanced quality of life and substantial capital accumulation, even if it results in wage adjustments due to automation.

AI Firms and Centralized Planning

The discussion touches on the potential for AI-driven firms to implement centralized planning, leveraging AI’s capabilities to optimize operations beyond human management.

Tameh Besiroglu ([175:42]) discusses how AI can enhance centralized decision-making by improving information processing and reducing principal-agent problems inherent in human-run organizations.

“If you can fine-tune AI systems to align with desired outcomes, it could dramatically change the structure of organizations, making centralized planning more efficient.”
— Tameh Besiroglu

Addressing Objections to Explosive Growth

Several objections to the notion of explosive economic growth driven by AI are addressed, including:

1. Bottlenecks and Scaling Limits: The guests argue that while bottlenecks exist, the overall increase in compute and complementary innovations will still drive significant growth.

2. Regulatory Constraints: They acknowledge potential regulatory hurdles but believe the economic incentives to adopt AI will often override these barriers.

3. Economic Distribution: Concerns about who benefits from rapid growth are discussed, with the guests positing that capital ownership and productivity gains will ensure broad economic benefits.

Tameh Besiroglu ([159:15]) responds to concerns about output distribution by asserting that even with bottlenecks, the reallocation of labor and capital will sustain high growth rates.

Final Thoughts and Advice for Interested Individuals

Concluding the episode, Ege and Tameh offer advice for those interested in contributing to AI development:

• Engage with Communities: Actively participate in communities and networks that focus on AI and related fields to stay informed and collaborate effectively.

• Prioritize Learning: Focus on key literature and research papers that shape the current understanding of AI advancements.

• Maintain Flexibility: Stay adaptable and ready to update one’s understanding as the AI landscape evolves rapidly.

Ege Erdil ([181:25]) emphasizes the importance of being part of a community to stay aligned with impactful research and innovations.

“Seek out people that have similar views and you're able to have very high bandwidth conversations with and seemingly make progress on these topics.”
— Ege Erdil ([181:25])

Conclusion

Ege Erdil and Tameh Besiroglu present a measured perspective on the trajectory towards AGI, emphasizing the complexities and multi-faceted nature of technological and economic growth. They challenge the notion of an imminent intelligence explosion, advocating for a broader understanding of how AI integrates with various economic sectors and the infrastructural demands required for sustained advancement. Their insights call for nuanced discussions and strategic planning to navigate the transformative potential of AI in the coming decades.

Notable Quotes:

• “It wasn't just that we had more horsepower... It was a bunch of complementary changes to many different sectors in the economy.”
  — Tameh Besiroglu ([00:50])

• “We're just getting a ton more compute every single year for the next few years... What is wrong with this logic?”
  — Keshav Murugesh ([45:05])

• “I think people underemphasize the support that is had from the overall upgrading of your technology of the supply chains.”
  — Tameh Besiroglu ([77:33])

• “People are going to have preferences that are different and less constrained by biology.”
  — Ege Erdil ([120:42])

These quotes encapsulate the essence of the discussion, highlighting the interplay between compute scaling, economic factors, and the multifaceted challenges in achieving AGI.