The LLM Revolution: From ChatGPT to Industry Adoption
Artificial Intelligence (AI) is currently experiencing a significant surge in popularity. Following the viral success of OpenAI’s conversational agent, ChatGPT, the tech industry has been abuzz with excitement about Large Language Models (LLMs), the technology that powers ChatGPT. Tech giants like Google, Meta, and Microsoft, along with well-funded startups such as Anthropic and Cohere, have all launched their own LLM products. Companies across various sectors are rushing to integrate LLMs into their services, with OpenAI counting customers like fintech companies using them for customer service chatbots, edtech platforms like Duolingo and Khan Academy for educational content generation, and even video game companies like Inworld for providing dynamic dialogue for non-playable characters (NPCs). With widespread adoption and a slew of partnerships, OpenAI is on track to achieve annual revenues exceeding one billion dollars.
While these results are impressive and might suggest the obsolescence of knowledge workers, there is a crucial distinction between LLMs like GPT-4 and human experts: LLMs lack genuine understanding. Their responses are not based on logical reasoning but rather on statistical operations. These models are trained on extensive internet data, including text from social media, wikis, forums, news, and entertainment websites, encompassing billions or trillions of words. To learn how to generate coherent text, LLMs train on millions of text completion examples. For instance, they might predict the next word after seeing phrases like “It was a dark and stormy night” or “The capital of Nigeria is Abuja.” Over time, they become adept at this task, especially when the next word is almost always the same, such as “The capital of Nigeria is.” However, in more ambiguous contexts like “It was a dark and,” LLMs may choose what a human would consider a reasonable option like “stormy,” but maybe “sinister” or “musty” instead. This initial phase of training is known as pretraining. To enhance their ability to perform specific tasks, like writing code or engaging in casual conversations, LLMs undergo further training on targeted datasets designed for those purposes.
However, LLMs’ training process, focused on predicting likely next words, can lead to a phenomenon known as hallucinations, where they confidently produce incorrect information and explanations when prompted. This happens because LLMs select the most plausible option when faced with multiple possible completions, even if that option lacks any basis in reality.
For example, Google’s chatbot, Bard, made a factual error during its first public demo when it stated that the James Webb Space Telescope took the first pictures of a planet outside our solar system. In reality, the Very Large Telescope had achieved this in 2004, while the JWST wasn’t launched until 2021.
Hallucinations are not the only shortcoming of LLMs. Training on massive amounts of internet data also leads to issues related to bias and copyright. Bias refers to the differing outputs from an LLM across attributes like race, gender, class, or religion. Since LLMs learn from internet data, they tend to inherit human biases and prejudices. Furthermore, training on vast datasets raises concerns about copyright issues, where the use of copyrighted materials in training LLMs without proper consent and compensation has led to legal disputes.
While efforts are underway to address these issues, existing LLMs are not infallible. Users must validate and review text generated by LLMs for accuracy, factuality, and potential biases. Remember that LLM-generated content should be treated as a first draft, not the final version. The responsibility for scrutiny and revision lies with the users to ensure the quality and integrity of the content.
