Merge pull request #25 from MIT-Emerging-Talent/gemma

Gemma

Author: AseelOmer

commercial_models/README.md

@@ -1,29 +1,51 @@
-# Commercial Models – Green AI Analysis
+# Commercial Models: The Crisis of AI Environmental Visibility
+
+The environmental impact of Large Language Models (LLMs) has transitioned from
+an esoteric concern to a critical research priority. While early studies focused
+primarily on the energy cost of the massive **training phase** (e.g., GPT-4
+training estimates: $\approx 51.8 \text{ – } 62.3 \text{ GWh}$) [Source: Sustainability
+by Numbers], recent, rigorous analyses—such as the landmark benchmarking paper
+*"How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM
+Inference"* (Jegham et al., 2025)—have confirmed two major points:
+
+* **Inference Dominance:** The energy consumed during **inference**
+* (running the model
+for every user query) is increasingly the dominant contributor to an LLM's total
+lifecycle environmental footprint due to global scale [Source: Google Cloud Blog].
+* **Divergent Reporting:** LLM energy reporting is highly unstable. As
+demonstrated by the benchmarking paper, resource consumption can vary by over
+**70 times** between the most and least efficient models for an equivalent task.
+Furthermore, consumption figures depend not just on model size, but profoundly
+on **deployment infrastructure** (PUE, regional carbon intensity) [Source: AWS
+Sustainability Report] and **methodology** (batch size, token length, hardware
+generation) [Source: Optimal Sparsity of MoE, arXiv]. This divergence leads to
+figures where one study reports less consumption (e.g., Epoch AI's $0.3 \text{ Wh}$
+per query) [Source: Epoch AI], while another (using different infrastructure
+assumptions) reports significantly more.
 
-This folder contains the **comparative sustainability analysis** of several
-large language models (LLMs) used in commercial applications.  
-It is part of the **ELO2 Green AI project**, focusing on estimating
-the **energy, carbon, and water footprints** of each model.
-
----
-
-## 📄 Contents
-
-- **models.md** – Main document providing technical summaries and
- sustainability estimates for:
-  - GPT-4 (OpenAI)
-  - Claude 3 Haiku (Anthropic)
-  - Gemini Nano (Google)
-  
 ---
 
-## 🎯 Purpose
-
-This documentation:
-
-- Highlights how **different LLM architectures and deployments**
-  affect energy and water use.
-- Demonstrates how **model size and hosting** influence environmental impact.
-- Supports ongoing evaluation of **Green AI strategies** for efficient computing.
-
----
+## Our Research Intervention
+
+This critical lack of standardized, reliable environmental data creates an
+**eco-efficiency paradox**: we cannot definitively prove which models are
+sustainable without standardized measurements, but the required performance
+trade-offs often push users toward the most power-hungry frontier models (e.g.,
+GPT-4's high capability, estimated at $0.3 \text{ – } 1.8 \text{ Wh}$ per query)
+[Source: The Verge].
+
+This project addresses this paradox with an **interventionist approach**.
+Rather than
+simply measuring existing systems, our research question directly investigates
+whether we can *engineer* open-source models to definitively close the gap in the
+efficiency balance:
+
+> **Research Question:** **Can open-source language models, enhanced with
+optimization techniques such as recursive editing and distillation, become
+environmentally and functionally viable alternatives to commercial models?**
+
+We hypothesize that by applying resource-saving techniques (e.g., **distillation**
+to compress model size and **recursive editing** to refine output with minimal
+re-computation), we can achieve an **eco-efficiency score** that consistently
+demonstrates their viability as a sustainable alternative, driving adoption of
+open-source models over proprietary, resource-intensive solutions.