Which is Worse, AI or Crypto?

So I got to wondering this week why I’m not hearing nearly as much about the environmental costs of Generative AI model training as I did about crypto mining. People were all over crypto miners about their environmental impact but, with a few minor exceptions, I haven’t heard this brought up as being an issue with Generative AI or AI research in general.

Sure, I could look up some studies, get some government info, find some tables, try to do a little bit of correlation and analysis but then I figured “Why not just launch a token?” No. I figured “Why not just ask Gippity?”

So here we go!

The Initial Prompt - Environmental Impact Comparison

“I’m curious about the environmental impact of training and operating Generative AI models versus mining cryptocurrency. Back when crypto was the big thing, lots of people were talking about the carbon footprint of mining operations based on the proof of work standard.

These days we hear a lot about the amount of compute required to train LLMs but I don’t hear much about the environmental impact. Can you compare and contrast the power usage of Generative AI training versus crypto mining and speculate on why the environmental cost of crypto was so frequently mentioned versus that of Generative AI training?”

Skip to the Bottom to Read the Whole Thing

I’ll include the entire transcript of the conversation below but the jist of GPT’s response was that crypto mining running full time used far more energy than ”one-shot” LLM training and that, due to it‘s speculative nature, crypto drew more criticism from environmentalists than even GenAI would at the same level of energy consumption.

I thought it curious that GPT addressed GAI training as though it only happened once… which I pushed back on a bit, suspicious that it was minimizing the energy consumption numbers by focusing only on the training of, say, an individual LLM rather than the huge investments in compute that are ongoing for Generative AI.

Second Prompt - What About Ongoing Training?

”While the one-time training phase may be relevant for a specific LLM, are we not continuously training new LLMs, and so does removing this factor appear to lessen the overall environmental impact of Generative AI development on a global scale?“

GPT admitted that the ongoing training of new models and updating of old ones should be factored into the global environmental footprint of Generative AI but then reiterated that negative public perception of AI development, even if its footprint is larger than we might expect, is mitigated by the perceived benefits of Generative AI and the fact that, unlike crypto mining, the work of developing it is spread across a wide variety of companies that people already trust and respect.

Nothing was said, of course, about the concerns regarding Generative AI training being based on publicly accessible data as this wasn’t specifically relevant to the environmental impact.

Third Prompt - Actual Numbers

”What are the actual energy consumption amounts per year at the peak of crypto mining versus the current energy draw of Generative AI training and application?”

ChatGPT reported that peak crypto mining consumed 100 to 150 terawatt hours of electricity per year, enough energy to power the entire nation of Argentina, whereas Generative AI training, so far, is consuming about 10 terawatt hours per year. It then went on to take a few shots at crypto‘s energy consumption for being centralized on two major blockchains (bitcoin and ethereum) as well as crypto being perceived as kind of shady in the first place.

Final Prompt - Checking The Conclusion

“So based on this information, one might simplify the argument by saying that not only did crypto mining consumed roughly an order of magnitude more power than Generative AI but also crypto itself was speculative and therefore more easily questionable in nature whereas we’re already seeing meaningful contributions across a wide range of business and individual applications for AI. Would you agree that this is an acceptable summary?

GPT agreed that crypto mining received more scrutiny because it consumed an order of magnitude more power, noted that crypto mining was speculative in nature and therefore more highly suspect and, and pointed out that Generative AI is already providing meaningful contributions in business and personal use… positive societal impact leading to lowered environmental scrutiny.

Changing Perception of Generative AI

One has to wonder if Generative AI will continue to get the environmental impact pass it’s currently enjoying should its energy consumption rise significantly or if, more likely, people decide that its development, should it fail to produce the Advanced General AI that we’re all being promised, has been a tech investor bubble all along.

I would anticipate significant growing concerns about the environmental impact of Generative AI in the future. It seems we’re happy to trash the planet so long as we feel we’re getting something we want out of it, but less so when it turns out it was just about putting money in someone else’s pocket all along.

And to those nay-sayers who claim that AI is going to destroy our society... well from where I stand, with our addiction to pump and dump economic cycles and flexible moral and ethical principles about environmental impact, not to mention the injustice of training GenAI on unlicensed and uncompensated content, it seems we’re perfectly capable of doing that to ourselves.

—————————————————

The Transcript in Full

Prompt 1

I’m curious about the environmental impact of training and operating Generative AI models versus mining cryptocurrency. Back when crypto was the big thing, lots of people were talking about the carbon footprint of mining operations based on the proof of work standard.

Response 1

The environmental impact of both cryptocurrency mining and training generative AI models is significant, but they differ in scale, visibility, and public perception.

Cryptocurrency Mining:

Energy Usage:

Proof of Work (PoW) Mechanism: Most cryptocurrencies, particularly Bitcoin, rely on the PoW mechanism, which requires solving complex mathematical puzzles to validate transactions. This process is highly energy-intensive.

Continuous Demand: Mining operations run 24/7, and as more miners join the network, the puzzles become more complex, increasing energy consumption.

Global Impact: As of recent years, Bitcoin alone was estimated to consume more electricity annually than some small countries, drawing significant public scrutiny.

Environmental Concerns:

Carbon Footprint: A large portion of crypto mining occurs in regions with high reliance on fossil fuels, leading to substantial carbon emissions.

Public Outcry: The direct association of mining with carbon emissions, combined with its speculative nature, led to a broad public discourse on its environmental harm.

Generative AI Training:

Energy Usage:

Training Phase: Training large language models (LLMs) like GPT involves processing vast amounts of data, requiring substantial computational resources over weeks or months. This consumes significant energy.

Operational Phase: Once trained, the energy required to operate and infer from these models is less than the training phase but still substantial, especially as models are deployed at scale.

Environmental Concerns:

Compute-Intensive: Training a large AI model can have a carbon footprint equivalent to the lifetime emissions of several cars. However, this is a one-time cost per model, compared to the continuous energy draw of crypto mining.

Location and Energy Source: The carbon footprint depends heavily on where the data centers are located and the energy mix of those regions. Many AI companies are moving towards renewable energy sources, which can mitigate some of the environmental impact.

Why Crypto's Environmental Impact Was More Publicized:

Public Perception:

Speculative Nature of Crypto: Cryptocurrency, particularly during its boom, was often seen as speculative, with critics questioning its societal value. The environmental cost added fuel to this debate, making it a focal point in discussions.

High Visibility: Crypto mining operations are often large-scale and concentrated, making them more visible targets for criticism.

Generative AI's Perception:

Utility and Innovation: AI is widely perceived as a transformative technology with applications across industries, from healthcare to education. This positive perception might overshadow environmental concerns.

Less Continuous Visibility: The environmental cost of AI training is not ongoing like crypto mining. Once a model is trained, the subsequent energy usage is much lower, making the issue less persistent in public discourse.

Conclusion:

In summary, while both crypto mining and AI training have substantial environmental impacts, the continuous and highly visible nature of crypto mining, combined with its speculative reputation, made it a more prominent target for environmental scrutiny. Generative AI, despite its high initial energy costs, benefits from a more positive societal perception and a one-time training phase, leading to less frequent discussions about its environmental footprint.

Prompt 2

Response 2

You're right—while each large language model (LLM) undergoes a one-time training phase, the ongoing development of new models and frequent updates to existing ones mean that the training process is continuous on a global scale. This continuous cycle of training new models, fine-tuning existing ones, and developing specialized versions contributes to the overall environmental impact of Generative AI.

Continuous Training and Its Implications:

New Models: As technology advances, new models are regularly developed to improve performance, incorporate more data, or add new capabilities. Each new model undergoes a training process that consumes significant computational power and energy.
Model Updates: Even after an LLM is initially trained, it may require periodic retraining or fine-tuning to incorporate new data, adjust to changing environments, or refine its performance. This retraining also consumes energy.
Specialized and Smaller Models: Beyond the major LLMs, many specialized or smaller models are trained for specific tasks or industries. Although these models may consume less energy individually, their collective impact can be substantial.
Global Impact: The combined effect of training and retraining across the AI industry means that the environmental impact is ongoing and potentially growing as the demand for more sophisticated models increases.

Comparative Environmental Impact:

Crypto Mining vs. Generative AI Training:

While crypto mining is continuously intensive due to the nature of Proof of Work, generative AI training represents a series of high-energy spikes rather than a continuous draw. However, the frequency and scale of these spikes across the industry make the cumulative energy consumption significant.

Both processes contribute to substantial energy usage and carbon emissions, but AI training is often distributed across various companies and research institutions, which can diffuse public attention.

Public Perception and Discourse:

Visibility: The continuous cycle of AI model training is less visible than the 24/7 operations of crypto mining, which may explain why its environmental impact has not received as much public scrutiny.

Perceived Value: AI's broad applicability and perceived positive societal contributions may also lead to less criticism compared to the more controversial and speculative nature of cryptocurrencies.

Conclusion:

In essence, while individual AI models have a one-time training cost, the industry as a whole is engaged in continuous training, which adds up to a substantial and ongoing environmental impact. This makes the distinction between "one-time" and "continuous" less clear when considering the global scale of AI development.

Prompt 3

”What are the actual energy consumption amounts per year at the peak of crypto mining versus the current energy draw of Generative AI training and application?”

Response 3

The energy consumption of cryptocurrency mining and Generative AI can be substantial, but they are difficult to compare directly due to differences in scale, transparency, and how the data is reported. However, we can look at estimates to get a sense of their respective impacts.

Cryptocurrency Mining Energy Consumption:

Bitcoin Mining:

Peak Energy Consumption: At its peak, Bitcoin mining consumed an estimated 100-150 terawatt-hours (TWh) of electricity annually. This was equivalent to the energy usage of a small country like Argentina.

Global Share: Bitcoin alone was responsible for around 0.5% of global electricity consumption at its peak.

Ethereum (Pre-Merge):

Ethereum's Consumption: Before Ethereum transitioned from Proof of Work (PoW) to Proof of Stake (PoS) in September 2022, it was estimated to consume around 50 TWh annually.

Generative AI Energy Consumption:

Training Large Language Models (LLMs):

Training Costs: The training of a single large AI model, such as GPT-3, was estimated to consume several gigawatt-hours (GWh) of electricity. For instance, GPT-3's training was estimated to consume around 1,287 MWh (1.287 GWh), which is roughly equivalent to the energy consumption of 120 average U.S. homes for a year.

Operational Costs:

Inference Costs: The energy required for deploying and running these models across millions of devices also adds up. While each individual query or inference may consume milliwatts or watts of power, the cumulative effect can be substantial.

Cumulative Impact: Given the rapid growth in the number of AI applications and models being trained and deployed, the overall energy consumption by the AI industry is increasing.

Comparative Summary:

Bitcoin Mining: At its peak, Bitcoin mining consumed 100-150 TWh per year.

Generative AI: The total energy consumption of Generative AI is harder to estimate precisely. However, the energy required for training individual large models (e.g., several GWh per model) and the operational costs across the industry are growing. While no precise annual total exists, some estimates suggest that the AI industry could consume tens of TWh annually, depending on the scale of operations.

Why Crypto's Energy Impact Was More Discussed:

Transparency: Cryptocurrency mining energy consumption is easier to estimate and track due to the transparency of the blockchain and the focus on a few major cryptocurrencies like Bitcoin and Ethereum (pre-Merge).
Concentration: Mining operations are often concentrated in specific geographic regions, making their impact more visible.
Public Perception: The speculative nature of cryptocurrency led to more public scrutiny, whereas AI is seen as having broader societal benefits, which may reduce the focus on its energy costs.

Conclusion:

In conclusion, while the energy consumption of Generative AI is significant and growing, the peak energy consumption of cryptocurrency mining, particularly Bitcoin, was more clearly defined and easier to communicate, leading to more public discussion about its environmental impact.

Prompt 4

Response 4

Yes, that's a well-phrased and accurate summary. It captures the key points:

Energy Consumption: Cryptocurrency mining, particularly at its peak, consumed roughly an order of magnitude more power than current Generative AI operations.
Nature of the Technology: Cryptocurrency was largely speculative in nature, leading to heightened public scrutiny over its environmental impact.
Perceived Value: In contrast, Generative AI is already providing meaningful contributions across various sectors, which has influenced public perception and led to less focus on its environmental costs.

This summary effectively conveys the core differences in energy consumption and societal impact between cryptocurrency mining and Generative AI.