Failing to Understand the Exponential, Again
8 hours ago
- #ai-progress
- #exponential-growth
- #future-predictions
- The article compares the current discourse around AI progress to the early weeks of the Covid-19 pandemic, where exponential trends were underestimated.
- People often dismiss AI's potential despite its rapid advancements, such as writing programs and designing websites, due to occasional mistakes.
- METR's study shows exponential improvement in AI's ability to complete long software engineering tasks, with recent models like GPT-5 performing tasks over 2 hours.
- OpenAI's GDPval study measures AI performance across 44 occupations, showing GPT-5 nearing human performance and Claude Opus 4.1 outperforming it in some areas.
- The article predicts significant AI advancements by 2026, including models working autonomously for full days and matching human experts across industries.
- Extrapolating current trends suggests AI will frequently outperform experts by 2027, with 2026 being a pivotal year for AI integration into the economy.
- The underperformance of models like Grok 4 and Gemini 2.5 Pro highlights the risks of overfitting to benchmarks (Goodhart's law).