The social cost of LLMs

I take great pride in my work – at least lately, if not permanently. I have information at my fingertips. I have creativity at my disposal. I have answers to as many questions as I have ever had since my birth. I LLM. The second-best part of my day is when I place myself behind closed doors and surrender to the magic that LLMs are. Which one is the best part? It depends on who is asking! 

For every “me”, there is a “him” who does not have access to this information. As if getting an internet connection was not tough enough, now we want him to get more computing muscle in the form of CPUs and GPUs. My colleague says LLMs are amplifiers. I agree, and I am sure they are. They increase the gap between the haves and have-nots, widening the digital divide. 

Energy requirements for training LLMs and using them for inferencing have skyrocketed, putting massive pressure on the electricity supply. We are running out of training data. Imagine this: Suddenly, the whole Internet is not enough data. Because we are flooding the internet with AI-generated content now, training datasets for building future LLMs will lack original human-generated content. The problem amplifies in cases of languages that are underrepresented online. 

We gain productivity in our daily lives and work. However, don’t get me started on the job displacements that will occur if workers’ time is not repurposed mindfully. There is a huge mismatch between the demand and supply of talent to work with generative AI (both building LLMs from scratch and AI Engineering, which is building applications using LLMs). We can discuss the Big AI Bang Theory more another time. I am not surprised that AI’s rapid evolution and its potential to revolutionise various fields mirror the cosmic expansion. 

Rapid progress has effectively concentrated economic power within a few countries with the wherewithal to research and build models. Online presence, or rather the lack of it, has handicapped specific languages in the minority. Underrepresentation in the training dataset is a significant technical minus. 

Remember the intellectual property (IP) ownership concerns and debates? Bias and a lack of fairness in the responses generated by LLMs have made the ethics guardians hypervigilant. The general public is worried about infringement of their privacy, and data breaches keep organisations on their toes. If hallucinations and the resulting misinformation are not handled well, recommendations from LLMs could be catastrophic. 

The rise of LLMs is a double-edged sword, offering immense potential while exposing profound societal challenges. Bridging the digital divide and ensuring equitable access must be prioritised to prevent further marginalisation. Collaboration across nations, industries, and communities is essential to balancing innovation with inclusivity and fairness. It is not enough to be born as humans. We must also ensure that we are human by the time we leave. 

The AI race is touted as the most significant in the past several decades and the decades to come. I say it is a rat race at the end of the day. The winner-takes-all approach leaves others as losers unless they start calling themselves learners.

Staring at death, I see / Other rats like me / Staring at death, I see / I want to be free.

Far away, I see / The starting line in front of me / Had I smiled more / Had I laughed more / Had I talked more / Had I sung more.

How unfair, I don’t get a second chance / How unfair I cannot start again / Had I lived more / Had I lived more / Had I lived more.