It is undeniable that ChatGPT and Generative AI with large language models (LLMs) have, in recent times, revolutionized many areas of computing and applications in life generally.
Whether it be coding, travel planning, language translation, content summarization, creating pictures or videos, other machine learning related applications, or CV enhancement, there is scarcely a field that has not been touched with the rise of generative AI, exemplified by ChatGPT. It has renewed the interest in artificial general intelligence (AGI) and has led to a spike in investments by all sorts of small and big companies who do not want to be left out. This is aside from companies like Microsoft, Google, Anthropic and OpenAI who are directly involved in the development of ChatGPT and other LLMs.
However, there are some issues with generative AI as well, that may prevent it from realizing its full potential or lead to its eventual fall from the current hype cycle. In this article we will discuss some of the issues that critics have raised.
The blog article by Paris Marx (https://disconnect.blog/the-chatgpt-revolution-is-another/) eloquently lists some of the issues.
- One of the issues is the amount of computing power consumed in training models like ChatGPT. This includes the direct costs, including cost of electricity and carbon footprint. There are huge costs of training the models and separate costs in inferencing the millions of requests. This can be several millions of dollars and as much as $1.2 billion, as per this link for ChatGPT5. A Forbes article mentions that ChatGPT consumes half a million kW of electricity daily which is equivalent to 180000 U.S. households. This has raised serious questions about the sustainability of ChatGPT and other LLMs. Another article highlights the environmental impact of ChatGPT and data centres in general.
- There are additional costs for the infrastructure such as the water for cooling the servers, as in this link. This is also clearly unsustainable.
- Another issue is the fact that the high costs of training such a model from scratch (despite the promising open source tools and frameworks like RAG and finetuning that try to make the costs of customizing a large language model for a specific use case more affordable) can only be afforded by a few big companies who can afford to invest in it, which then get to choose what algorithms are used to train it and other important parameters. This is clearly not equitable or democratic.
- Despite all the hype about ChatGPT and allied models, the capabilities and reliability of such models for real life applications have often been questioned. For example, the problem of hallucinations, where the generative AI model generates wrong answers with a tone of absolute confidence, has not been completely solved yet.
- Philosophers and scholars like Noam Chomsky have questioned whether we are any close to real intelligence or general intelligence as a result of these chatbots, and psotulated that a different kind of model that is closer to how the human brain works can do the job better.
- There is the general fear that ChatGPT will perpetuate all that is wrong with AI and make the problems worse in terms of employment (in areas such as coding or content writing), AI led surveillance and existing inequalities in this world.
- The way in which the ChatGPT model (and all such models typically) is trained is by using cheap labour in developing countries to tag the responses and the inputs. This too perpetuates existing inequalities.
- The silicon valley relies on a constant stream of such hype cycles to continue the high investment that is poured into it, without showing real value, and the ChatGPT is merely the latest of these hypes and one that will fizzle away in some time.
- Among the multiple open source LLMs that are available, generally speaking paid models such as ChatGPT (which has a cost for using their APIs in terms of number of tokens used) and others like Claude outperform the open source models. Added to that is the cost of hosting the models, with users having to pay for AWS or other infrastructure for developing a real life end to end application. This makes them out of reach for small companies or developers who do not have sufficient financial backing for their projects.
So, to conclude, while ChatGPT does have many promising use cases and its widespread use is increasing, there are still many open questions and it is not a given that they will widely revolutionize industry and society. Only time can tell how useful they will eventually be, or whether this is only the latest example of the hypes that seem to come up from time to time.
