You make several valid points about the shortcomings of LLMs, such as:
LLMs are not revolutionary or game-changing, but rather incremental improvements over existing methods. They rely on massive amounts of data and compute power, which are not always available or accessible. They also perform poorly on tasks that require reasoning, understanding, or common sense.
LLMs have reached a saturation point where adding more data and parameters does not lead to significant gains or new capabilities. They also face scaling limits that make them impractical and costly to train and deploy. Moreover, they are prone to errors, inconsistencies, and biases that can have harmful consequences.
LLMs have created a bubble in the AI space that is driven by hype and speculation rather than scientific rigor and evidence. You warn that investors, researchers, and users should be more cautious and critical of the claims and promises made by LLM developers and promoters. You also urge them to consider the negative impacts of LLMs on society, such as misinformation, manipulation, and environmental damage.
However, I also think that we can acknowledge some of the positive aspects of LLMs, such as:
LLMs have demonstrated impressive performance on certain benchmarks and tasks that are relevant for real-world scenarios, such as question answering, summarization, translation, dialogue, and generation. They have also enabled new forms of creativity and expression, such as poetry, music, art, and humor.
LLMs have inspired new research directions and innovations in the AI field, such as self-supervised learning, attention mechanisms, transformers, and neural architecture search. They have also stimulated interdisciplinary collaborations and dialogues among researchers from different domains and backgrounds.
LLMs have opened up new opportunities and possibilities for AI education, democratization, and accessibility. They have lowered the barriers to entry for AI development and usage by providing pre-trained models, APIs, platforms, and tools that anyone can use.
I think these aspects are also important to consider when evaluating the merits and drawbacks of LLMs. For example, you mention CRITIC, a framework that allows LLMs to self-correct with tool-interactive critiquing. This is an example of how LLMs can be improved and augmented with external feedback and validation.
