Issues¶
Environmental Issues¶
Sources of carbon footprint
- train models
- data collection
- inference
- creating new chips
Types¶
- Embodied carbon
- Operational carbon
Not just the energy cost of running computers, but also energy required to build them
Solutions¶
- Efficiency of models
- Software infra to promote development of efficiency
- AutoML instead of grid search
- Easy-to-use APIs for efficient models
- Efficient computer chips
- Efficient datacenters
- Microsoft: underwater servers
- 8 x more reliable, using Nitrogen, etc
- No need for fresh water for cooling
- Use renewable energy
- Transitioning to 100% renewables will not eliminate the carbon footprint of chips
- Stop planned obscelence
4 M’s of AI Efficiency¶
- Model
- Machine
- Mechanism
- Map: Location
Jevon’s Paradox/Effect¶
Improved efficiency in resource utilization increases the total consumption of resources, due to increased rate of consumption from increased demand
Hence, definition of “efficiency” is important
PUE¶
Power Usage Effectiveness
Lower is better
Typically \(\in [1.1, 1.6]\)
How the power input is being used for compute & for other supporting consumption such as cooling, lighting, etc. $$ \begin{aligned} \text{PUE} &= \dfrac{\text{Total Energy}}{\text{Productive Energy}} \ &= 1 + \dfrac{\text{Overhead}}{\text{Productive Energy}} \end{aligned} $$
Ethical Issues¶
Abstraction¶
Researchers/Engineers often abstracted away from application they’re working on, hence not aware of bad things that can be used for
Data Ownership¶
Data governance
Trade-off: Privacy vs Progress
Model monetization trained on consumer data, eg: GitHub Copilot
Data Bias¶
Dataset curation
Federated learning
Research Inequality¶
- Inequitable access to computing resources
- Inequitable access to datasets