llm/b9e3557d-356c-4503-a149-187e2ce27f4b/ee99f7e8-5ce4-40d8-af1f-b9e6d787a069-output.json
# Summary: Early Indicators for AI Progress and Impact This Notion page documents an ongoing collaborative research project focused on identifying measurable early indicators for AI progress and its potential societal impacts. The project aims to develop metrics that could provide advance warning of significant AI capabilities or economic disruption. ## Project Approach and Methodology The project leader has found that serious work on complex forecasting questions requires extended multi-turn conversations rather than group discussions. Success has come through persistent 1:1 engagement via email and calls, though some collaborative workshops (like Helen's CSET workshop) have worked for brainstorming and aggregating existing ideas. The plan involves multiple phases: initial deep analysis through individual conversations, followed by group brainstorming sessions to generate concrete measurement proposals. ## Key Research Questions (Cruxes) The project examines several crucial uncertainties about AI development: - When will "superhuman coders" and AI researchers emerge, as defined in AI 2027 scenarios? - How is AI task completion horizon progressing? Research from METR shows AI time horizons (tasks AI can complete with 50% success rate) have been doubling approximately every 4-7 months since 2019. - What explains the gap between benchmark performance and real-world task completion? Current frontier models show ~50-minute time horizons on software tasks but struggle with reliability and long-horizon planning. - Are certain skills emerging as bottlenecks for automation, particularly "research taste" and judgment-laden tasks? - Can algorithmic improvements decouple AI impact from compute investment? ## Economic and Adoption Considerations Significant attention is given to whether current AI investment levels constitute a bubble. Evidence suggests AI adoption is occurring faster than almost any technology in history—ChatGPT reached 100 million users within two months. However, questions remain about whether this translates to productivity gains. One analysis found no evidence of increased software shipping rates despite widespread AI coding tool adoption, with developers potentially overestimating their productivity improvements by 20% or more. The project notes concerns about circular economics in AI investment, where AI companies are major customers of each other's services. OpenAI projects unprecedented revenue growth from $10B to $100B over three years, which would be historically anomalous. ## Capability Gaps and Bear Case Arguments A detailed "bear case" argues current LLM approaches may face fundamental limitations. Key observations include: - LLMs excel at "in-distribution" problems present in training data but struggle with genuine out-of-distribution generalization - The "jagged frontier" of capabilities—where models excel at some tasks while failing at similar ones—suggests memorization rather than true intelligence - Real-world agents remain fragile, with reliability issues persisting despite benchmark improvements - Test-time compute and reinforcement learning may not generalize beyond domains with easy verification ## Measuring Real-World Impact The project emphasizes that many crucial human capabilities remain unmeasured by existing benchmarks: managing complexity across long projects, metacognition and dynamic planning, continuous learning and memory, leveraging external information, judgment in ambiguous situations, and genuine creativity. These capabilities are essential for tasks like building businesses, planning activities, or providing professional services. The research suggests tracking sector-specific AI diffusion as leading indicators, with software engineering potentially serving as a "fast example" while identifying slower-adopting sectors for comparison. Collaborators are exploring partnerships with organizations like Epoch AI, METR, and government agencies to develop comprehensive measurement frameworks.