Will AI completely replace software engineers according to the talk?

No. He argues that even when AI handles most tasks, engineers will shift to higher-level work such as management and system design. The idea is about role evolution, not immediate obsolescence. (Timestamp: 70)

What does he say about the diffusion speed of AI progress?

He presents two poles: one where progress is slow and diffuses gradually, and another where progress is rapid via recursive self-improvement. He notes real-world momentum, such as rapid revenue growth at Anthropic, suggests diffusion can be fast in practice. (Timestamp: 166)

Is there evidence of a software renaissance in practice?

He questions whether the world shows a renaissance of software features, noting that even with cloud code, we haven't clearly seen broad new features that wouldn't exist otherwise. This leads to a nuanced view of impact. (Timestamp: 119)

What revenue growth figures does he cite for Anthropic?

He mentions a sequence: zero to $100 million in 2023, $100 million to $1 billion in 2024, and roughly $1 billion to $9–$10 billion in 2025, illustrating explosive growth. (Timestamp: 216)

What book/topic does he reference related to technology's trajectory?

He references 'Adolescence of Technology' as a context for discussing the spectrum of technological progress. (Timestamp: 86)

Where does he hint that gains might come from beyond coding itself?

He notes that even with AI handling much coding, there are gains from higher-level activities and system management, implying effects beyond just code generation. (Timestamp: 75)

Dario Amodei On How His AI Predictions Have Been Misunderstood

The Lunar Society

Science & Technology3 min read4 min video

Feb 20, 2026|3,253 views|59|22

Save to Pod

Key Moments

TL;DR

AI will write most code fast; jobs evolve, not disappear.

Key Insights

AI predictions about coding unfold along a spectrum (90% of code, 100% of code, 90% of end-to-end SWE tasks), not a single binary milestone.

Even with high AI productivity, human engineers shift to higher-level work (design, coordination, architecture) rather than being replaced.

There’s potential for significant demand reduction in SWE jobs (up to 90%), but new roles and leadership opportunities emerge for humans.

Adoption can be fast in practice, contrasting with views that AI progress diffuses slowly; real-world business momentum can accelerate rapidly.

Economic signals (e.g., Anthropic’s revenue growth) provide tangible evidence that AI productivity translates into real business value, though broad systemic gains may take time.

Closing the loop on self-contained systems and broader software gains remains challenging, yet momentum suggests a material software renaissance is underway.

THE SPECTRUM OF AI CODING IMPACT

At the core, Amodei argues that AI's impact on software unfolds along a spectrum, not a single moment. He distinguishes benchmarks like 90% of code written by models, 100% of code, and 90% of end-to-end SWE tasks (including compiling, environments, testing, and even memos). He stresses these are different milestones, so earlier fears of a world with no engineers are misplaced. The takeaway is that progress will be rapid but with layered, variant targets rather than an all-or-nothing jump.

FROM 90% TO 100%: WHAT CHANGES AND WHAT DOESN'T

Moving from 90% to 100% of code or tasks is not just a global productivity boost; it's a qualitative shift in what humans do. Even if models handle most coding, humans will take on higher-level control, architecture, design, and coordination work. He notes that even when 90% of tasks are automated, 10% of critical decisions still require human judgment, and the dynamic of collaboration changes as engineers focus on system-level thinking rather than routine implementation. The practical implication is routine re-skilling and new leadership roles, not mass layoffs.

ECONOMIC IMPACT AND THE ENGINEERING WORKFORCE

Beyond individual projects, Amodei frames economic impact in terms of value capture and labor mix. The potential for a 90% reduction in SWE demand exists, but so do opportunities for engineers to tackle higher-value tasks. The 'adolescence of technology' analogy to farming emphasizes that disruption comes in stages: early wins scale quickly, but broad transformations take time as ecosystems mature. The takeaway is that productivity gains may outpace job destruction, with new roles driving innovation and reliability work that only humans can perform.

DIFFUSION SPEED, RECURSION, AND THE TWO POLARS

He contrasts two opposite narratives about change speed: a slow diffusion story and an explosive self-improvement story. He cites Anthropic's revenue growth—zero to 100M in 2023, 100M to a billion in 2024, and to roughly 9-10B in 2025—as evidence that AI-enabled capabilities can scale quickly in practice, despite concerns about diffusion. This section emphasizes that real-world adoption often defies simple timing models, and technological progress can be punctuated by rapid market responses even amid organizational complexity.

REAL-WORLD SIGNALS: REVENUE MOMENTUM AND THE BROADER PICTURE

Finally, Amodei reflects on broader software development dynamics beyond the pure coding tasks. He notes that greenfield cloud projects and the overall software renaissance remain uneven and highly context-dependent, with many projects starting slowly. Yet the repeated pattern of revenue acceleration signals that AI-driven productivity translates into tangible business outcomes, supporting a case for continued investment. The discussion ends by acknowledging that while closing the loop on self-contained systems remains a challenge, the momentum is undeniable and continuing.

Mentioned in This Episode

●Books

Common Questions

He outlines a spectrum with milestones like 90% of code written by AI, moving toward 100% code written by AI, and then 90% of end-to-end SWE tasks being AI-written. He emphasizes that these benchmarks are distinct and progress is accelerating, not uniform. (Timestamp: 43)