Capabilities · Trend
Science of AI Training Dynamics
New mechanistic analyses of AdamW training dynamics and Weibull weight distributions advance the scientific understanding of how transformer models evolve during training.
Forecast horizon
Connections
Connections · 9
How this node ties into the rest of the map, and the evidence behind each link.
to · enables 4/10
Zero-Knowledge Proof Verification of AI Training Compute A scientific understanding of training dynamics is a prerequisite for designing verifiable training specifications used in ZK proof architectures.
+4 growth from · informs 4/10
LLM Probabilistic Reasoning Limitations Systematic probabilistic reasoning failures in LLMs motivate studying training dynamics to understand why these biases emerge and persist.
+4 growth from · informs 4/10
Weibull Weight-Scale Dynamics in AdamW Training New result links pretraining data predictability to Weibull weight-scale growth, advancing science of training dynamics.
+4 growth from · drives 3/10
Subquadratic LLM Architecture Breakthrough Claims Claims of a subquadratic architecture breakthrough would fundamentally alter the science of AI training dynamics and scaling laws.
+3 growth from · informs 3/10
Subquadratic LLM Architecture Breakthrough Claims Subquadratic architecture claims, if verified, would represent a fundamental shift in LLM training dynamics.
+3 growth from · supports 3/10
Scaling Laws for LLM Social Simulation Fidelity Scaling law analysis for social simulation fidelity contributes to the broader science of AI training dynamics.
+3 growth from · drives 3/10
AI for Science Needs Reasoning, Not Just Data Public argument that scientific AI must prioritize reasoning machinery rather than data scale alone.
+3 growth from · constrains 3/10
Limits on Near-Term AI Recursive Self-Improvement Skeptical RSI timelines emphasize practical training and oversight bottlenecks over unconstrained self-improvement.
+3 growth from · informs 2/10
Scaling Laws for LLM Social Simulation Fidelity Scaling law analysis of social simulation fidelity contributes to the science of AI training dynamics.
+2 growthSignal sources
Signal sources
Dated facts from primary sources in this direction.
Task horizon doubling Mar 2025
The length of software tasks AI agents can do autonomously at 50% reliability has doubled about every 7 months — and since 2024 closer to every ~3 months.
METR → Benchmarks saturating Apr 2025
In one year scores rose by 18.8, 48.9 and 67.3 points on MMMU, GPQA and SWE-bench; real-world software solve rate jumped from 4.4% to 71.7%.
Stanford HAI — AI Index 2025 → Autonomous coding 2025–2026
On SWE-bench Verified (500 real GitHub issues), autonomous coding agents reached ~80–86% by late 2025, up from under 50% in early 2025.
Epoch AI →