[HUMAN VOICE] "AI Timelines" by habryka, Daniel Kokotajlo, Ajeya Cotra, Ege Erdil
Nov 17, 2023
auto_awesome
Ajeya Cotra, Daniel Kokotajlo, and Ege Erdil, researchers in the field of AI, discuss their varying estimates for the development of transformative AI and explore their disagreements. They delve into concrete AGI milestones, discuss the challenges of LLM product development, and debate factors that influence AI timelines. They also examine the progression of AI models, the potential of AI technology, and the timeline for achieving super intelligent AGI.
The experts strongly disagree on the timeline for transformative AI, with estimates ranging from four to 40 years for when 99% of currently remote jobs become automatable.
The relative importance of compute capabilities and other factors contributes to their differing timelines.
The complexity of implementing useful AI systems and uncertainties surrounding factors like in-context learning and adversarial robustness make AI forecasting challenging and drive disparate timelines.
Deep dives
Differing Timelines for Transformative AI
Ajayakotra, Daniel Cockatailo, and Ege Adil have spent considerable time investigating the timeline for transformative AI. Despite their research, they strongly disagree on the relevant timescales. For example, their median estimates for when 99% of currently remote jobs become automatable range from four to 40 years. Graphs illustrating their differences show probability distributions that diverge significantly. The conversation explores the reasons behind their disagreements, considering factors such as compute overhang, in-context learning, adversarial robustness, and the role of government and companies in either slowing down or accelerating AI development.
Compute Overhang and AGI Timelines
A potential crux of disagreement between Daniel and Ajayakotra revolves around whether transformative AI will be achieved before or after the current compute overhang is exhausted. Daniel assigns a higher probability to a substantial R&D speed-up before the overhang is depleted, leading to shorter timelines. Ajayakotra's timelines, on the other hand, are influenced by small changes in compute requirements, as this pushes a significant probability mass into a pre-overhang world. The relative importance of compute capabilities and other factors play a role in their differing timelines.
Kinks and Detail in AI Development
A recurring theme in the discussion is the recognition that AI development involves working out kinks and dealing with the complexity of implementing systems that are actually useful. Ajayakotra and Daniel acknowledge the need for time and effort to overcome challenges in architecture, training setup, and other areas. While Daniel believes that additional AI engineers will significantly speed up AI R&D, Ajayakotra is more cautious, expecting a longer timeframe due to the intricacies involved in making AI systems equivalent to current human engineers.
Uncertainties and Meta-Level Disagreements
There are additional uncertainties and meta-level disagreements throughout the conversation. Factors such as in-context learning, adversarial robustness, market efficiency, and the role of government regulations and company attitudes shape the varying viewpoints. These uncertainties make it challenging to assign precise probabilities and lead to disagreements even on small updates. The conversation highlights the complex nature of AI forecasting and the difficulty of pinpointing specific factors that drive disparate timelines.
AI Progress and AGI Possibilities
The podcast episode explores different scenarios and potential timelines for AGI development. The first main idea discussed is the possibility of AGI happening in the near future and its potential implications. The second main idea focuses on the potential capabilities and limitations of AI systems by 2030, including their impact on AI research and the economy.
The Path to AGI and Transformative AI Systems
The podcast episode delves into a hypothetical scenario that outlines the progression of AI systems leading up to AGI. The scenario highlights notable milestones, such as the development of multi-modal models and their applications in various tasks. It also discusses the increasing autonomy and productivity of AI systems, their potential role in AI research and development, and the uncertainties surrounding the timeline and impact of AGI.
How many years will pass before transformative AI is built? Three people who have thought about this question a lot are Ajeya Cotra from Open Philanthropy, Daniel Kokotajlo from OpenAI and Ege Erdil from Epoch. Despite each spending at least hundreds of hours investigating this question, they still still disagree substantially about the relevant timescales. For instance, here are their median timelines for one operationalization of transformative AI:
