The New Scaling Law
1. Key Themes
The Shift from Pre-Training to Continual On-the-Job Learning
The dominant AI scaling paradigm — spend more on pre-training, get a smarter model — is giving way to a new one: models that learn continuously through feedback and iteration during deployment.
"AI has been defined by pre-training: spend more upfront, get a smarter model. But a shift is underway, from pre-training to continual, on-the-job learning."
On-the-Job Learning Velocity Is Itself Accelerating — and Compounding
This isn't just a new capability; the rate of improvement in that capability is itself rapidly improving, suggesting a compounding dynamic that could dramatically separate frontier models from laggards.
"The newest models learn on the job almost twice as fast as models from just three months earlier, and the gains compound with iteration."
EdgeBench Data Confirms the Trend Is Systematic, Not Anecdotal
The attached chart (EdgeBench score gains over 2 hours, Aug 2025–Jun 2026, log scale) shows a clear upward trend across multiple model families — OpenAI's GPT-5 Codex series, Anthropic's Claude Opus series, GLM-5, and DeepSeek-V4-Pro — all trending upward along a consistent regression line. GPT-5-Codex began near a score gain of ~2.5; by May 2026, GPT-5.5 reached ~17, and Claude Opus 4.6 briefly exceeded 16. This is not a single-vendor phenomenon; it is market-wide.
2. Contrarian Perspectives
The next moat in AI may not be who has the biggest pre-training budget, but who has the best feedback loops. The conventional wisdom among investors and observers has been that AI dominance is a function of compute spend and data at training time. This article challenges that directly:
"A shift is underway, from pre-training to continual, on-the-job learning."
If on-the-job learning velocity is the new scaling law, then companies with the richest real-world feedback environments (e.g., those with high-volume, high-stakes human interactions) may compound their model quality advantage faster than those with larger upfront budgets but fewer deployment feedback loops.
Models doubling their learning speed every ~3 months implies capability timelines are being compressed faster than most forecasts assume. The EdgeBench chart shows near-doubling of score gains from Aug 2025 (GPT-5-Codex, ~2.5) to roughly Jan 2026 (Claude Opus 4.6, ~16) — a 6x gain in about 5 months on a log scale. If this trend continues, the practical capability ceiling for deployed AI agents could arrive significantly earlier than consensus projections.
3. Companies Identified
OpenAI
- Description: AI lab, creator of the GPT model family
- Why mentioned: Multiple GPT-5 Codex variants (GPT-5-Codex, GPT-5.1-Codex, GPT-5.2-Codex, GPT-5.3-Codex, GPT-5.4, GPT-5.5) plotted on the EdgeBench chart, showing consistent improvement in on-the-job learning speed from Aug 2025 through May 2026
- Quote (chart label): GPT-5.5 reaches an EdgeBench score gain of ~17 by May 2026, up from ~2.5 for GPT-5-Codex in Aug 2025
Anthropic
- Description: AI safety-focused lab, creator of the Claude model family
- Why mentioned: Claude Opus 4.5 through 4.8 plotted on EdgeBench; Claude Opus 4.6 briefly led the benchmark in Jan 2026 with a score gain exceeding 16
- Quote (chart label): Claude Opus 4.6 peaks above 16 on the EdgeBench score gain axis in Jan 2026
Zhipu AI (GLM)
- Description: Chinese AI lab, creator of the GLM model family
- Why mentioned: GLM-5 and GLM-5.1 appear on the EdgeBench chart, demonstrating that the on-the-job learning scaling trend is not exclusive to US frontier labs
- Quote (chart label): GLM-5.1 achieves a score gain of ~8 by Mar 2026
DeepSeek
- Description: Chinese AI research company
- Why mentioned: DeepSeek-V4-Pro appears on the EdgeBench chart in Mar 2026, though it scores below the trend line (~5), suggesting it currently lags in on-the-job learning velocity relative to frontier peers
- Quote (chart label): DeepSeek-V4-Pro plotted at approximately 5 on the EdgeBench score gain axis
EdgeBench
- Description: AI benchmarking platform measuring on-the-job learning speed
- Why mentioned: Primary data source for the new scaling law thesis; measures average EdgeBench score gain over 2 hours across model families
- Quote: "Source: EdgeBench. Y-axis in logarithmic scale."
4. People Identified
Max (last name not provided)
- Description: Analyst or partner at Coatue Management
- Why mentioned: Presents the C:\Takes video segment elaborating on the new scaling law thesis
- Quote: "For more on this C:\Take, watch Max"
5. Operating Insights
Design products and workflows that maximize model feedback loops, not just model quality at launch. If on-the-job learning velocity is the new performance driver, then the product or enterprise that exposes AI to more high-quality, iterative feedback during use will see compounding capability gains that a static integration will not.
"The newest models learn on the job almost twice as fast as models from just three months earlier, and the gains compound with iteration."
Re-evaluate AI vendor selection criteria. Procurement and build-vs-buy decisions historically optimized for benchmark scores at release. The new frame should weight a model's learning velocity in your specific environment — i.e., how quickly it improves given your data and feedback — over its initial capability score.
"A new study measured whether models improve when given time, feedback, and room to work."
6. Overlooked Insights
The logarithmic scale of the EdgeBench chart means the real-world performance gap between top and bottom performers is far larger than the chart visually suggests. DeepSeek-V4-Pro and GLM-5 appear close to Claude Opus 4.6 and GPT-5.5 on the chart, but because the Y-axis is log-scale, the actual score gain difference is multiplicative — frontier models may be 3–6x more capable at on-the-job learning than they visually appear relative to laggards. Investors and operators should not underestimate this gap.
"Source: EdgeBench. Y-axis in logarithmic scale."
The ~10-month window captured (Aug 2025–Jun 2026) suggests this data is forward-looking or projected, meaning Coatue may be publishing thesis-driving data before it is widely available to the market — a potential early signal for investors tracking which model families will dominate agentic deployments in H2 2026.