AGI is still 30 years away — Ege Erdil & Tamay Besiroglu - Dwarkesh Podcast Recap

Podcast: Dwarkesh Podcast

Published: 2025-04-17

Duration: 3 hr 8 min

Summary

Ege Erdil and Tamay Besiroglu argue that while AI is progressing rapidly, achieving full automation of remote work tasks may still be decades away due to the complexity of required capabilities and limitations in scaling compute power. They emphasize that the perception of imminent AGI is often overly optimistic.

What Happened

In this episode, hosts Ege Erdil and Tamay Besiroglu, co-founders of Mechanize, delve into the future of artificial intelligence and the timeline for achieving artificial general intelligence (AGI). Tamay challenges the notion of an intelligence explosion by likening it to calling the Industrial Revolution a 'horsepower explosion'. He argues that the broader changes in society and technology were the result of various complementary innovations, not just a single factor like horsepower. This perspective invites listeners to rethink how they view advancements in AI, suggesting that while smart AI systems will emerge, they are just one component of a much larger transformation.

The discussion also touches upon the timeline for remote worker automation, with Tamay suggesting a timeline around 2045 for full automation of tasks that can currently be done remotely. Ege, on the other hand, is more optimistic, predicting a shorter timeline but acknowledges that significant hurdles remain. They note that while recent advancements in AI have been impressive, the actual fraction of the economy that has been automated remains small, leading them to believe that the road to full automation is longer than many might expect. They highlight the need for further breakthroughs in AI capabilities and the challenges of scaling compute resources necessary for these advancements.

Key Insights

Key Questions Answered

What is Tamay Besiroglu's perspective on the intelligence explosion?

Tamay Besiroglu argues that the concept of an intelligence explosion is not particularly useful. He compares it to labeling the Industrial Revolution as a 'horsepower explosion', suggesting that such a view overlooks the multifaceted innovations that contributed to the era's transformation. He emphasizes that while advancements in AI will lead to smarter systems, this is just one part of a broader array of changes that will drive economic and technological organization.

When do Ege Erdil and Tamay Besiroglu expect full automation of remote work?

Ege Erdil and Tamay Besiroglu offer differing timelines for when full automation of remote work might be achieved. Tamay suggests around 2045 for complete automation, while Ege is a bit more optimistic, indicating a shorter timeframe. However, both agree that achieving full automation is not just about speed but also about the complexity of tasks involved and the current limitations of AI.

How much of the economy has currently been automated by AI?

The hosts highlight that despite the rapid advancements in AI technology, the fraction of the economy that has actually been automated remains quite small. They point out that many people extrapolate current trends to predict imminent AGI or widespread automation, which can be misleading. This underscores the importance of understanding the limitations of current AI capabilities and the challenges in scaling those capabilities across various sectors.

What breakthroughs are necessary for achieving AGI according to the hosts?

Ege and Tamay discuss the need for significant breakthroughs in AI competencies to achieve AGI. They identify that while there have been substantial advancements, such as improvements in reasoning and coding, there are still many core capabilities that AI systems need to develop further. This includes achieving coherence over long timeframes and full multimodal understanding, akin to human capabilities.

What challenges do the hosts identify in scaling AI compute power?

The discussion reveals that significant challenges lie in scaling AI compute power. They mention that the previous decade saw a substantial increase in compute capabilities, but as we approach the limits of scaling, there are concerns about energy and resource constraints. Ege notes that sustaining this scaling may require a considerable fraction of global output, which raises questions about the economic viability of such a path.