LessWrong (Curated & Popular)

“My AGI timeline updates from GPT-5 (and 2025 so far)” by ryan_greenblatt

Aug 21, 2025
The discussion delves into the rapid advancements in AI, especially with GPT-5's capabilities. It explores the state of software engineering tasks and how performance metrics have improved. The timeline for complete automation of AI research is scrutinized, revealing a doubling time for progress that's faster than before. While there's excitement around these developments, the host reflects on the nuances of this growth and the challenges that still lie ahead in fully realizing AI's potential.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

2025 Showed Modest Acceleration

  • AI progress in 2025 was somewhat faster than historical rates but lacked large jumps.
  • Doubling time for horizon length improved from ~210 days historically to ~135 days in 2025, showing modest acceleration.
INSIGHT

One-Time Bump Versus Sustained Trend

  • Recent gains may be driven by a one-time release bump with subsequent progress near historical pace.
  • Pre-training returns appear weaker, making RL scale-ups relatively more influential in 2025.
INSIGHT

Limits Of Horizon-Length Benchmarks

  • Horizon lengths beyond about one month become hard to measure and less meaningful.
  • Super-exponential progress may matter more at longer horizons and could accelerate later on.
Get the Snipd Podcast app to discover more snips from this episode
Get the app