Lens 1: Stochastic Parrots

Scale and Environment

Moritz Mähr

moritz.maehr@gmail.com

University of Bern

What did BLOOM’s training cost?

Documented case

Which costs disappear when we count only the training power?

Estimates for final training: Luccioni et al. (2023), Table 3. Original chart, CC BY 4.0.

Start with the chart. Participants have read Bender et al.; do not start with another summary of the paper. Ask them to identify the parts left out of the first bar. The expanded boundary includes 14.6 t from idle infrastructure and 11.2 t from hardware manufacture. Rounded values sum to 50.49 t; the paper reports 50.5 t. These are estimates built from measured energy and accounting assumptions, not a complete footprint for every use of BLOOM (Luccioni et al. 2023).

BLOOM has 176 billion parameters. It was trained on Jean Zay in France. Its ROOTS training corpus has 46 natural and 13 programming languages (Laurençon et al. 2022; Luccioni et al. 2023). The example makes Bender et al.’s environmental costs concrete (Bender et al. 2021, sec. 3).

The model runs somewhere

Hardware · electricity · cooling · location

Site context photographed in 2026: Clemcloum8, CC0 (Clemcloum8 2026). The panel is not evidence for BLOOM’s training footprint.

The panel makes the physical infrastructure visible. It was photographed after BLOOM’s training. Do not subtract its heat-recovery claim from the 2023 paper’s estimates. The photo does not establish the hardware configuration or heating arrangements during BLOOM’s training.

The learning question is: Which environmental and data choices are hidden by a model’s apparent performance? Bender et al. connect environmental costs with unequal distributions of benefit and harm (Bender et al. 2021, sec. 3). Crawford follows this material base through mines, grids, and supply chains (Crawford 2021, ch. 1).

Optional background belongs here, not on the projected slide: the IEA’s data-center totals are sector-wide estimates and projections (International Energy Agency 2025). The water figure in Li et al. is an estimate with explicit assumptions (Li et al. 2025). Neither provides a per-request measurement for the group’s own tool.

Training data: who survives the filter?

Hypothetical illustration of exclusion. Concept: Bender et al. (2021), sec. 4. Original diagram, CC BY 4.0.

Ask which page the rule loses. The pages and the specific rule are invented for teaching. Do not present them as a sample from C4 or BLOOM.

Then connect the example to the documented argument. Bender et al. discuss how C4’s blocklist filtering excludes LGBTQ-related text and how web-scale collection overrepresents dominant viewpoints (Bender et al. 2021, sec. 4). A larger dataset does not establish broader representation. The choice of a filter is a value choice. The filtering example is a separate illustration; it does not describe BLOOM’s data pipeline.

Documentation debt: what can you trace?

Source

Where did the training text come from?

Filter

What was excluded, and by whom?

Record

Can someone check those decisions?

An undocumented choice becomes harder to question.

Concept: Bender et al. (2021), sec. 4.

Recall documentation debt from the assigned paper: collecting more data than one can document makes later accountability harder (Bender et al. 2021, 615). Use the fictional filtering example: if the filter’s rule is missing, how would the excluded group identify what happened?

BLOOM’s open reporting illustrates what can be recorded. ROOTS and the BigScience governance work document data sources and decisions (Laurençon et al. 2022; Jernite et al. 2022). This does not make BLOOM cost-free or settle consent for all data. Ask participants which records exist for the tool they use.

A training footprint does not price your task

Final training

50.5 t CO₂eq

One accounting boundary.

One deployment

914 kWh over 18 days

One observed pattern of use.

For your task: what still needs measurement?

Separate studies within Luccioni et al. (2023). The values have different units and boundaries.

Do not add the two numbers. The 18-day API deployment served 230,768 requests; its hardware and conditions differ from final training. The study reports that keeping the model ready accounts for much of deployment energy (Luccioni et al. 2023, sec. 4.4). It does not give the energy for a different research workflow.

Luccioni et al. compare generative and task-specific systems for several tasks (Luccioni et al. 2024). A search tool, manual work, or a smaller model is an alternative to test at comparable quality. Do not promise a saving without checking the task and the boundary.

Worked map: two stages are enough

Hypothetical worked analysis

System: BLOOM. Selected stages: Training; Deployment and use.

Stage	Known	Unknown	Avoidable (PROPOSED)
Training	Final training: 50.5 t CO₂eq	Cost of training for our task	Test an existing model before new training
Deployment and use	One 18-day deployment: 914 kWh	Energy for our intended use	Compare fewer requests and a task-specific tool

Largest unknown: Energy for our task. One avoidable cost: A new training run, if an existing model meets the need.

Plenary sentence: “Performance hides the cost of our intended use.”

Figures: Luccioni et al. (2023). The research team is fictional. Actions are proposals with unmeasured savings. Other stages: Not examined.

Show every filled field. This is a fictional research team considering BLOOM, using documented measurements. The known figures come from the paper. The unknowns concern the team’s intended use. The actions are proposed choices, not measured savings.

Mark all other stages Not examined. Explain that a blank cell is ambiguous; write Unknown when the evidence is missing. The Lens 1 group can use this prepared case if a Research Use cannot be discussed openly.

Takeaway

A missing number is an unknown, not a zero.

Ask about environmental costs, training data, and documentation debt.

Return to Bender et al. (2021), secs. 3–4.

Keep the lens slot at 15 minutes: case first, 5 minutes; concepts and worked map, 10 minutes. The Lens 1 group applies the four guiding questions later in the morning. Transition to Lens 2: We followed data as a cost. Next we follow data as a picture of the world.

Sources

• Bender et al. (2021): environmental costs, training data, documentation debt.
• Luccioni, Viguier, and Ligozat (2023): BLOOM footprint and deployment study.
• Laurençon et al. (2022); Jernite et al. (2022): ROOTS and data governance.
• Luccioni, Jernite, and Strubell (2024): comparing deployment tasks.
• Photo: Clemcloum8 (2026), CC0. Original diagrams: CC BY 4.0.

Full references and evidence limits