Four Time Scales for Technology Development and Deployment

4 hours ago

Technology development occurs across four distinct time scales: new research ideas (10-20+ years), hype generation (short-lived cycles), at-scale deployment (software ~20 years, hardware longer), and reshaping the economy (50+ years).
New research ideas often take decades to mature, with many false starts; for example, neural networks evolved from 1943 to 2012 for deep learning, and another decade for LLMs.
Hype cycles rapidly inflate expectations but rarely lead to transformative technologies, as seen with blockchain, metaverse, IBM Watson, and nanotechnology; many people confuse hype with real progress.
Scaling a technology from solid product to mass adoption requires huge effort; Unix/Linux took over 40 years to dominate, and self-driving cars are still not widely deployed despite decades of work.
Reshaping the economy takes over 50 years of continuous deployment, as with electrification or containerization, yet hypesters often claim AI and robotics will transform the world in years.
The main mistake is confusing early research or hype with imminent, large-scale impact, leading to overly optimistic and sometimes damaging predictions.

Intelligence alone is often not the main bottleneck in the real world; regulation and governance are key constraints.
In medicine, clinical trials consume seven years and over a billion dollars per drug, slowing progress despite AI advances.
Biomedical innovation is hindered by factors like target herding and patent systems that disincentivize novel biology.

The AI as Normal Technology framework is useful for understanding AI's economic impact unless a future discontinuity like recursive self-improvement occurs.
Recursive self-improvement (RSI) is unlikely to suddenly cause mass unemployment, as economic transformation from AI will be gradual over decades.
Future jobs will change significantly, shifting from executing tasks to evaluating and controlling AI systems, emphasizing judgment and domain knowledge.

Critics of AI spending are judging models trained on outdated hardware, ignoring the upcoming massive wave of compute.
Global AI chip capacity is set to increase 6x between 2024 and 2026, based on TSMC CoWoS wafer allocations.
Nvidia leads AI silicon supply, with Broadcom (Google TPUs) and AMD also significantly increasing production.

The author criticizes the excessive hype around AI, labeling it 'AI Confidence Theater' that causes more harm than good.
Many claimed 'life-changing' AI use cases are actually basic workflows, and true transformative applications are rare.
This hype distorts reality, hurts innovation, and leads to toxic comparisons where people feel demoralized and behind.

A platform shift occurs roughly every 10 years, similar to AWS's cloud-native shift in 2006, with AI (especially LLMs) potentially causing another fundamental change.
AI provides immediate value like cloud did, with easy-to-use HTTP APIs and solutions for formerly hard problems (e.g., sentiment analysis, code generation), but this value may be mistaken for durable value.
Early AI is impractical for many contexts due to missing tooling and reliability issues, reminiscent of early cloud's limitations that shrank over time as capabilities expanded.

Intelligence alone is often not the main bottleneck in the real world; regulation and governance are key constraints.
In medicine, clinical trials consume seven years and over a billion dollars per drug, slowing progress despite AI advances.
Biomedical innovation is hindered by factors like target herding and patent systems that disincentivize novel biology.

The AI as Normal Technology framework is useful for understanding AI's economic impact unless a future discontinuity like recursive self-improvement occurs.
Recursive self-improvement (RSI) is unlikely to suddenly cause mass unemployment, as economic transformation from AI will be gradual over decades.
Future jobs will change significantly, shifting from executing tasks to evaluating and controlling AI systems, emphasizing judgment and domain knowledge.

Critics of AI spending are judging models trained on outdated hardware, ignoring the upcoming massive wave of compute.
Global AI chip capacity is set to increase 6x between 2024 and 2026, based on TSMC CoWoS wafer allocations.
Nvidia leads AI silicon supply, with Broadcom (Google TPUs) and AMD also significantly increasing production.

The author criticizes the excessive hype around AI, labeling it 'AI Confidence Theater' that causes more harm than good.
Many claimed 'life-changing' AI use cases are actually basic workflows, and true transformative applications are rare.
This hype distorts reality, hurts innovation, and leads to toxic comparisons where people feel demoralized and behind.

A platform shift occurs roughly every 10 years, similar to AWS's cloud-native shift in 2006, with AI (especially LLMs) potentially causing another fundamental change.
AI provides immediate value like cloud did, with easy-to-use HTTP APIs and solutions for formerly hard problems (e.g., sentiment analysis, code generation), but this value may be mistaken for durable value.
Early AI is impractical for many contexts due to missing tooling and reliability issues, reminiscent of early cloud's limitations that shrank over time as capabilities expanded.

Hasty Briefsbeta