Failing to Understand the Exponential, Again By Julian Schrittwieser, Sat 27 September 2025 The current debate around AI progress and claims of an AI “bubble” echoes early responses to the Covid-19 pandemic, where exponential trends were ignored. Similarly, skepticism about AI’s capability growth—due to occasional mistakes or apparent plateauing—misses the continual exponential progress being made. --- Exponential AI Progress Evidence: METR Study METR specializes in studying AI capabilities, measuring how long AI models can autonomously perform software engineering tasks. Clear exponential trends: Sonnet 3.7 completes tasks up to 1 hour at 50% success. Recent models like Grok 4, Opus 4.1, and GPT-5 exceed previous predictions, performing tasks of over 2 hours. The doubling rate of task length capability is roughly 7 months. METR’s updated graphs show continued upward trend despite public skepticism. --- Broader Economic Impact: OpenAI’s GDPval GDPval evaluates AI performance across 44 occupations in 9 industries, based on 1320 tasks rated against experienced professionals. Results show similar exponential improvements: GPT-5 is close to human-level performance in many tasks. Claude Opus 4.1 outperforms GPT-5 and nearly matches industry experts. OpenAI’s transparent inclusion of other labs' models (e.g., Claude Opus 4.1) promotes integrity. Be cautious about models claiming state-of-the-art results on narrow benchmarks (Goodhart’s Law). --- Outlook for AI Integration Exponential progress unlikely to stop suddenly; conservative extrapolations predict: Mid-2026: AI models able to work autonomously for full 8-hour workdays. End of 2026: At least one AI model matches human expert performance across many industries. End of 2027: AI frequently outperforms experts on numerous tasks. Simple linear extrapolation on well-measured metrics may better predict reality than many expert opinions. Recommended further reading: Epoch AI’s 2030 report AI 2027 project for a detailed future outlook. --- Tags AI, Progress, AGI, Safety, Productivity --- Footnotes: Note on models Grok 4 and Gemini 2.5 Pro underperforming on GDPval despite state-of-the-art claims, illustrating dangers of overfitting benchmarks (Goodhart’s Law). --- This article stresses the importance of understanding exponential progress in AI and warns against complacency based on superficial observations or limited benchmarks. The data from multiple rigorous studies paints a clear picture of rapid AI capability growth and impending widespread economic impact.