Dario Amodei on AI Progress: "We Are Near the End of the Exponential"
In a deep-dive conversation on the Dwarkesh Podcast, Anthropic CEO Dario Amodei discusses the limits of scaling laws, the future of reinforcement learning, and the economic integration of AI.
In a highly anticipated, long-form discussion on the Dwarkesh Podcast, Anthropic CEO and co-founder Dario Amodei sat down to dissect the current state of artificial intelligence development, model scaling, and the geopolitical landscape of computing.
As frontier labs push toward artificial general intelligence (AGI), Amodei offers a nuanced perspective: while the era of easy scaling via raw web data may be slowing down, new paradigms are opening up.
Q: You have suggested that we are approaching "the end of the exponential" in terms of traditional pre-training. What does that mean for the next generation of models?
Dario Amodei: The scaling hypothesis—the idea that adding more compute and data to standard transformers yields linearly better intelligence—is hitting a bottleneck. We have largely saturated the high-quality text available on the open internet.
However, that does not mean AI progress is stopping. We are transitioning from the "pre-training" era to the "post-training" era. The next exponential leap is coming from Reinforcement Learning (RL) and search during inference. Instead of just predicting the next word, we are teaching models to spend compute dynamically—running search trees and verifying answers before they respond. The scaling is shifting from training time to test-time.
Q: Anthropic has been vocal about security and data siphoning. How is the rise of automated model copying affecting safety?
Dario Amodei: We are seeing a major rise in what we call industrial-scale distillation. Companies are using API access to run millions of queries against our frontier models, effectively siphoning the reasoning pathways of models like Claude or Fable to train their own systems.
This presents a double danger. First, it undercuts the massive financial investments required to train base models. Second, it allows actors to copy safety-aligned systems and strip away their guardrails, distributing dangerous capabilities without the safety filters we spent months calibrating.
Q: How do you see the timeline to AGI, and what will the transition look like?
Dario Amodei: I still think we could see systems that resemble AGI by the end of this decade—possibly as early as 2027 or 2028. But the real bottleneck is no longer just the intelligence of the model; it is the diffusion rate in the physical economy.
Having an AI that is smart enough to be a lawyer or an engineer is different from integrating that AI into real-world legal systems or manufacturing pipelines. The transition will feel slower than many expect because human institutions, regulations, and safety protocols take time to adapt.