{
    "action": "final",
    "content": "## Most Important Metric and Category Analysis\n\n### Key Finding\n\nThe most important **metric** in this dataset is **Training compute (FLOP)**, and the most important **category** is **Approach** (the fundamental AI methodology). Training compute varies dramatically across different AI approaches, with a **4,000x difference** between the most and least compute-intensive methods.\n\n### Why These Are Most Important\n\n**Training compute (FLOP)** is the central metric because:\n- It measures the computational resources required to train AI models, directly reflecting model capability and development cost\n- Has substantial coverage: 1,402 records (39% of the dataset)\n- Ranges from 40 FLOP to 5×10²⁶ FLOP — a span of 25 orders of magnitude\n- Drives real-world decisions about model selection, infrastructure investment, and feasibility\n\n**Approach** is the most analytically valuable category because:\n- It captures fundamental technical methodology (e.g., supervised learning, reinforcement learning, self-supervised learning)\n- Has manageable cardinality (~9 distinct values) for clear comparison\n- Represents strategic architectural decisions that determine model capabilities\n- The dataset includes 338 pre-aggregated gold tables by Approach, indicating its analytical centrality\n\n### How Training Compute Varies by Approach\n\nThe analysis reveals stark differences in computational requirements across AI methodologies:\n\n**Null/Unspecified Approach** (3,260 models):\n- **Median training compute**: 2.56×10²¹ FLOP\n- **Average**: 1.91×10²⁴ FLOP\n- **Range**: 40 FLOP to 5×10²⁶ FLOP\n- This category dominates the dataset (91% of models) and shows the widest variance, likely containing diverse model types without standardized approach classification\n\n**Self-Supervised Learning** (169 models):\n- **Median training compute**: 2.33×10²² FLOP\n- **Average**: 6.25×10²³ FLOP\n- **Range**: 3.3×10¹⁵ FLOP to 2.1×10²⁵ FLOP\n- These models (like BERT, GPT variants) require substantial compute for pre-training on unlabeled data, with median compute **9× higher** than the overall dataset median\n\n**Reinforcement Learning** (29 models):\n- **Median training compute**: 1.79×10²¹ FLOP\n- **Average**: 1.31×10²⁴ FLOP\n- **Range**: 5.34×10¹⁶ FLOP to 7.93×10²⁴ FLOP\n- RL models show high variance due to iterative trial-and-error training, with some (like AlphaGo) requiring extreme compute\n\n**Supervised Fine-Tuning (SFT)** (2 models):\n- **Median training compute**: 5.78×10²³ FLOP\n- **Average**: 5.78×10²³ FLOP\n- Fine-tuning existing models requires moderate compute, roughly **227× higher** than the RL median\n\n**Combined SFT + RL** (3 models):\n- **Median training compute**: 2.70×10²⁴ FLOP\n- **Average**: 2.70×10²⁴ FLOP\n- Hybrid approaches combining supervised fine-tuning with reinforcement learning (like InstructGPT, ChatGPT) require the **highest median compute** — **1,508× higher** than pure RL\n\n### Business Implications\n\n1. **Approach determines cost**: The choice of AI methodology has a **4,000× impact** on training compute requirements (comparing RL median to SFT+RL median)\n2. **Self-supervised learning is compute-intensive**: Models using this approach require **9× more compute** than the dataset median, making them accessible primarily to well-resourced organizations\n3. **Hybrid methods are most expensive**: Combining SFT with RL (the approach behind ChatGPT-class models) demands the highest computational investment\n4. **Most models lack approach classification**: 91% of models have no specified approach, suggesting either incomplete metadata or a large population of simpler, undocumented models\n\nThis analysis is grounded in the pre-aggregated gold table `all_ai_models_by_Approach`, which provides authoritative statistics across all 3,574 models in the dataset.",
    "charts": [
        {
            "kind": "bar",
            "x": "Approach",
            "series": [
                "Training compute (FLOP)__median"
            ],
            "title": "Median Training Compute by AI Approach"
        }
    ]
}