LLMs are not a panacea for all problems

ChatGPT, and Large Language Models (LLMs) in general, took our world by storm. They have emerged as powerful tools in natural language processing, demonstrating remarkable capabilities in understanding and generating human-like text. ?

From maintaining an apparently intelligent conversation on any and all subjects, through being able to write works, reports, posts, more or less complex texts in any style, to writing code in any programming language to solve any problem, this technology seemed and still appears to be able to disrupt almost any job, educational process or any task at once. These models, such as GPT, have been trained on vast amounts of diverse data, enabling them to generate contextually relevant responses and assist with a wide array of tasks.

The current or future impact promises of this technology made it “the” technology topic of 2023 and likely beyond. I just received my usual newsletter from McKinsey and it contains three selected stories: “Generative AI in operations: capturing the value”, “The state of AI in 2023: Generative AI’s breakout year” and “Reset and reimagine: The role of generative AI in SG&A”.

However, beyond the more than natural excitement with this technology, it is crucial to recognize that the effectiveness of these sophisticated language models can be compromised under certain circumstances.

One of the primary and most obvious limitations of LLMs lies in their dependency on the data they’ve been exposed to during training. The model's performance is directly correlated with the quality and diversity of the training data, and it may struggle with tasks that fall outside the scope of its training set.

This latest article from HBR sheds some light on this topic Is GenAI’s Impact on Productivity Overblown? (hbr.org)

For instance, as we tested ChatGPT, we all faced questions or tasks related to a niche field or a topic with limited online representation, where the response was highly inaccurate or incomplete. In such scenarios, the model's lack of exposure to specialized information becomes apparent, emphasizing the importance of understanding the boundaries of its knowledge.

As the HBR’s article put it very clearly, “it is important to appreciate that LLMs’ leaps in syntactic fluency don’t translate into being better able to automatically look up facts”.

So that’s a bit of a problem, and should worry each and every user of LLM’s: the technology is capable of producing factual errors in an extremely convincing way.

Furthermore, productivity tends to decrease when LLMs are employed in tasks that demand reasoning or problem-solving beyond the scope of typical language patterns found online. While these models excel at generating coherent and contextually relevant text based on existing patterns, they may falter when confronted with new scenarios that require nuanced understanding or creative problem-solving.

The inherent limitations of these models become evident when faced with complex decision-making processes that demand more than the regurgitation of existing information. In such cases, relying solely on LLMs may hinder productivity and lead to suboptimal outcomes.

So, after the excitement, it is becoming increasingly clear that LLMs are a technology capable of explosively increasing the productivity of countless tasks. It is for current times like the internet or Google for previous generations.

But it’s not a panacea for all problems. And like HBR suggests “Leaders should consider where this technology actually helps and resist the urge to integrate it into every job and task throughout the organization. To do this, they need to understand two core problems of LLMs that are critical to their medium- and long-term business implications: 1) Its persistent ability to produce convincing falsities and 2) the likely long-term negative effects of using LLMs on employees and internal processes.”

For all who are not highly technical data science or LLM experts, and may want to know a bit more about the topic, I recommend the following reading Inside ChatGPT's Brain (serokell.io)