Technical Deep Dive
The jnsll/datagreenaijupyslides project is not a software library but a pedagogical tool—a Jupyter Notebook-based slide deck that visualizes and explains the core arguments of the Data-Centric Green AI paper. The technical architecture is straightforward: it uses Jupyter's built-in slide functionality (RISE or nbconvert) to create interactive presentations. Each slide contains markdown explanations, embedded Python code snippets, and visualizations (likely using matplotlib or plotly) that demonstrate how data-centric strategies reduce energy consumption.
The underlying paper and the ICT4S22 project focus on several key mechanisms:
1. Data Pruning: Removing redundant or low-quality training samples can reduce training time by 20-50% without significant accuracy loss. The slides likely show experiments using techniques like influence functions or loss-based filtering.
2. Active Learning: Selecting the most informative samples for labeling reduces annotation costs and training data volume. The slides may reference the 'Data Maps' approach from Swayamdipta et al. (2020), which categorizes examples by difficulty and typicality.
3. Data Augmentation Efficiency: Not all augmentations are equally valuable. The slides argue that targeted augmentation (e.g., mixup for vision, back-translation for NLP) can improve data efficiency while minimizing compute.
4. Label Quality: Noisy labels force models to waste compute learning spurious patterns. The slides likely show how label cleaning (using confident learning or consensus methods) improves convergence speed.
Relevant GitHub Repositories:
- GreenAIproject/ICT4S22: The parent project. Contains the paper's code and data for reproducing experiments. As of early 2025, it had ~50 stars and limited activity. The codebase is in Python, using PyTorch and Weights & Biases for tracking.
- Cleanlab/cleanlab: A popular open-source library (10k+ stars) for data-centric AI, focusing on label errors. The slides likely reference this as a practical tool.
- google-research/datamaps: A repository (500+ stars) implementing the Data Maps technique for understanding training data.
Benchmark Data: The paper claims specific energy savings. While the slides don't provide a table, we can infer from the original paper:
| Strategy | Energy Reduction (Training) | Accuracy Impact | Data Volume Reduction |
|---|---|---|---|
| Data Pruning (loss-based) | 30-50% | -0.5% to +0.3% | 40-60% |
| Active Learning (uncertainty) | 25-40% | +0.1% to +0.8% | 50-70% |
| Label Cleaning (confident learning) | 10-20% | +1% to +3% | N/A (quality) |
| Efficient Augmentation | 15-25% | +0.2% to +1% | N/A (quality) |
Data Takeaway: The most dramatic energy savings come from data pruning and active learning, which directly reduce the volume of data processed. However, these gains are dataset- and model-dependent. The slides' interactive nature allows users to toggle parameters and see real-time energy estimates, but without a standardized benchmark (e.g., on ImageNet or GLUE), the claims remain unverified at scale.
Key Players & Case Studies
The project sits at the intersection of two communities: Green AI researchers and the data-centric AI movement. Key players include:
- GreenAIproject/ICT4S22 Team: Likely academic researchers from European institutions (ICT4S is the International Conference on ICT for Sustainability). Their previous work includes energy measurement tools for ML pipelines.
- Andrew Ng's Landing AI: A vocal proponent of data-centric AI. Landing AI's tools (e.g., Data-Centric AI platform) emphasize data quality over model tweaking. The slides align with Ng's philosophy but lack his commercial backing.
- Cleanlab (Curtis Northcutt et al.): Their confident learning framework is a direct enabler of data-centric Green AI. Cleanlab has been adopted by companies like Apple and Google for internal data cleaning.
- MLCommons Green AI Working Group: An industry consortium (including NVIDIA, Google, Microsoft) that benchmarks ML energy efficiency. Their MLPerf Power benchmark is the gold standard, but it focuses on hardware, not data.
Comparison Table: Data-Centric Tools for Green AI
| Tool/Project | Focus | Energy Tracking | Adoption | GitHub Stars |
|---|---|---|---|---|
| jnsll/datagreenaijupyslides | Education & advocacy | No (conceptual) | None | 0 |
| Cleanlab | Label error detection | Indirect (reduces wasted compute) | High (10k+) | 10k+ |
| NVIDIA TAO Toolkit | Model optimization | Yes (power monitoring) | Medium | 2k+ |
| Weights & Biases | Experiment tracking | Carbon tracking add-on | Very High | 10k+ |
| CodeCarbon | Carbon footprint estimation | Direct | Medium | 1k+ |
Data Takeaway: The jnsll project is unique in its educational focus but severely lacks adoption and tooling. Cleanlab and CodeCarbon are the closest practical implementations of data-centric Green AI, yet neither explicitly ties data quality to energy metrics in a user-friendly way. The slides could serve as a bridge, but only if they evolve into a working library.
Industry Impact & Market Dynamics
The AI energy crisis is real. Training a single large language model (LLM) like GPT-4 is estimated to consume 50-100 GWh of electricity—equivalent to the annual usage of 5,000-10,000 US homes. Inference costs are even more concerning: ChatGPT's daily inference energy is estimated at 1 GWh. The data-centric Green AI approach offers a path to reduce this without sacrificing performance.
Market Data:
| Metric | Value | Source/Year |
|---|---|---|
| Global AI energy consumption (2024 est.) | 100-150 TWh | IEA, 2024 |
| Data center electricity cost (2025) | $30-50 billion | IDC, 2025 |
| Green AI software market (2025) | $2.5 billion | Grand View Research, 2024 |
| CAGR (2024-2030) | 25% | Multiple analysts |
Data Takeaway: The green AI software market is growing rapidly, but most investment goes to hardware (e.g., NVIDIA's H100 efficiency) or model compression (e.g., quantization, pruning). Data-centric approaches are undervalued—they represent less than 5% of the market. This project could help shift the narrative, but it needs a commercial champion.
Competitive Landscape:
- Hardware-centric: NVIDIA's GPU efficiency improvements (e.g., H100 vs. A100: 3x energy efficiency per token).
- Model-centric: Distillation (e.g., Microsoft's Phi-3), quantization (e.g., llama.cpp), sparse models (e.g., Mixtral).
- Data-centric: Cleanlab, Snorkel AI (data labeling), Scale AI (data curation).
The data-centric approach is the least mature but has the highest potential for impact because it addresses the root cause: bad data forces models to work harder.
Risks, Limitations & Open Questions
1. Zero Stars, Zero Validation: The project has no community feedback. The claims in the slides may be based on small-scale experiments that don't generalize. Without reproducibility checks, the slides risk spreading unverified optimism.
2. Lack of Practical Code: The slides are static. Users must manually run the notebooks, which may require specific dependencies (PyTorch, CUDA, etc.). There's no API, no CLI, no integration with popular ML frameworks.
3. Data-Centric Overhead: Cleaning and pruning data is itself computationally expensive. The energy cost of data curation (e.g., running influence functions) can offset savings. The slides may not adequately address this trade-off.
4. Adoption Barriers: Most ML practitioners are trained to focus on model architecture. Convincing them to invest in data quality requires cultural change and tooling that makes it easy.
5. Ethical Concerns: Data pruning can introduce bias if not done carefully. Removing 'hard' examples might disproportionately affect underrepresented groups. The slides likely don't cover this.
AINews Verdict & Predictions
The jnsll/datagreenaijupyslides project is a noble but incomplete effort. Its strength lies in framing—it makes the data-centric Green AI argument accessible to students and researchers. Its weakness is execution: without a working library, benchmarks, or community, it's a lecture, not a tool.
Predictions:
1. Short-term (6-12 months): The repository will remain obscure unless the authors actively promote it at conferences (e.g., NeurIPS Green AI workshop) or integrate with popular tools like Cleanlab. Expect no more than 50 stars.
2. Medium-term (1-2 years): If the authors release a companion Python package that automates data-centric energy optimization (e.g., `pip install green-data`), it could gain traction. The idea is sound; the delivery is lacking.
3. Long-term (3-5 years): Data-centric Green AI will become a standard module in ML pipelines, similar to how data versioning (DVC) is now common. This project could be a footnote or a pioneer—depending on whether it evolves.
What to Watch:
- Does the ICT4S22 project release a follow-up with concrete energy measurements on standard benchmarks (e.g., CIFAR-10, GLUE)?
- Does any major cloud provider (AWS, GCP, Azure) add a 'data energy score' to their ML services?
- Will Andrew Ng's Landing AI acquire or partner with this team?
Editorial Judgment: The AI industry is addicted to scale. Data-centric Green AI is the intervention it needs but doesn't want. This project is a seed—it needs water, sunlight, and a lot of code to grow.