The Old Paradigm vs. The New Paradigm
Old Paradigm: More Compute = Better Models
The traditional approach to AI improvement:
- More GPUs
- More Data
- More Parameters (θn+1)
The problem: Diminishing returns. Pre-training scaling is hitting limits.
New Paradigm: Compute + Signal = Better Models
The emerging approach requires both:
- Compute: Raw processing power
- Signal: Quality training data from RL environments
The Dual Bottleneck
AI progress is now constrained by both compute — as explored in the economics of AI compute infrastructure — for training AND quality environments to train on.
Key insight from Epoch AI: “Without diverse, high-quality environments, throwing more compute at RL risks wasting much of it.”
The Numbers That Matter
- ~$2.4K – Compute per RL task (cheap tasks = wasted compute)
- $1B+ – Anthropic discussed for RL environment spending
- $19B – OpenAI R&D compute (2026 projected)
The Strategic Implication
Quality training signals are now as important as raw compute.
The companies that can produce robust, diverse, reward-hack-resistant environments at scale may become as strategically important as chip suppliers.
This is part of a comprehensive analysis. Read the full analysis on The Business Engineer.
Frequently Asked Questions
What is The Shift from Compute to Signal: AI's New Paradigm?
The problem: Diminishing returns. Pre-training scaling is hitting limits.
What is the dual bottleneck?
AI progress is now constrained by both compute for training AND quality environments to train on.
What is the numbers that matter?
~$2.4K – Compute per RL task (cheap tasks = wasted compute). $1B+ – Anthropic discussed for RL environment spending. $19B – OpenAI R&D compute (2026 projected)
What is the strategic implication?
Quality training signals are now as important as raw compute.
More Resources
