Generative AI: what is it good for?

Historical Drivers of Advancement

• The 2017 introduction of the "Transformer" architecture by Google researchers created the "T" in GPT, enabling longer, more coherent text and code generation.
• The 2022 launch of GPT-3.5 as a public chatbot triggered the fastest adoption of consumer technology in history, reaching 100 million users within two months.

Technical Capabilities and Strengths

• Large language models process hundreds of billions of words of unlabeled internet data, eliminating the historical requirement for human-labeled datasets.
• Systems excel at pattern matching, style transfer, and generating convincing text across diverse personas and historical contexts.
• Generative AI has demonstrated proficiency in passing standardized professional exams, including the U.S. medical licensing and legal tests.
• Code generation offers a tight feedback loop where compilers or interpreters immediately flag errors, allowing for rapid correction.

Identified Weaknesses and Limitations

• The technology functions as a "black box" with over 100 billion weights that are too complex for humans to fully interpret or understand.
• Models lack transparency and are unreliable for tasks requiring the discovery of new facts or high-stakes accuracy.
• The British government and intelligence services have noted that these systems cannot yet be trusted for fact-finding without human verification.

Economic Impact and Workforce Projections

• OpenAI economists estimate 20% of the U.S. workforce could have 50% of their daily tasks affected by generative AI in the near future.
• Economic theory suggests that only full automation (100%) of a process can drive an "intelligence explosion" or exponential growth; partial automation leaves the human element as a rate-determining bottleneck.
• Current deployment will likely continue as a human-AI collaborative workflow rather than immediate full replacement of human roles.

Future Outlook and Risks

• Significant economic activity and innovation rates are expected to be altered, though progress pace may remain linear rather than exponential due to human oversight requirements.
• Full automation of processes is viewed as a prerequisite for achieving super-intelligence or runaway economic growth scenarios.