Demystifying LLM: Ensuring Transparency and Understandability

Pradeep Pujari
3 min readJun 2, 2024

LLM should not be a black box

Nearly two years after ChatGPT brought AI, particularly Large Language Models (LLMs), into the public spotlight, the primary focus in tech circles has been on the race to develop the necessary capabilities. A simple definition of a Large Language Model (LLM) is that it functions as a multi-task next-word generator.

To earn the public’s trust, large language models (LLMs) must provide accurate and hallucination-free responses. If responses require review and correction, it defeats the purpose and erodes public confidence. It should not function as a mere black box; it should also provide the reasoning behind why it considers the response correct. An index developed at Stanford University reveals the transparency of AI models. The AI Index report tracks, collates, distills, and visualizes data related to artificial intelligence (AI). Considering the potential dangers of AI, it is crucial for these companies to return to their more open and transparent practices. Transparency in AI encompasses two primary areas: inputs and models. Large language models, which serve as the foundation for generative AI, are trained by sourcing from internet to analyze and learn from diverse datasets, ranging from Reddit forums to Picasso paintings. In the early days of AI, researchers frequently disclosed their training data in scientific journals, enabling others to identify flaws by assessing the quality of the inputs.

Credit: Copied from Google Image Search

Today, key players often withhold the details of their data to protect against copyright infringement lawsuits and maintain a competitive advantage. This practice makes it difficult to assess the accuracy of AI-generated responses.

The models also lack transparency. How a model interprets its inputs and generates language depends on its design. AI firms often regard their model architecture as proprietary, seeing it as their “secret sauce”: the ingenuity of generative AI hinges on the quality of its computation. In the past, AI researchers released papers on their designs, but the race for market share has curtailed such disclosures. Without understanding how a model functions, it is difficult to evaluate an AI’s outputs, limitations, and biases. This opacity makes it challenging for the public and regulators to assess AI safety and guard against potential harms. This is particularly concerning. Google’s A.I. Search Errors Cause a Furor Online last week.

Google immediately fixed this issue though

Google’s recent unveiling of its new artificial intelligence (AI) capabilities for search, has sparked controversy due to a series of errors as shown above. More:

credit: From X (formerly twitter) platform

Governments have started to lay the foundation for AI regulations. While welcome, these measures focus on guardrails and “safety tests” rather than full transparency. The reality is that technologies are evolving too quickly. Therefore, regulators should call for model and input transparency, and experts at these companies need to collaborate with regulators.

AI has the potential to transform the world for the better — perhaps with even more potency and speed than the internet revolution. Companies may argue that transparency requirements will slow innovation and dull their competitive edge, but recent AI history suggests otherwise. These technologies have advanced through collaboration and shared research. Returning to those norms would increase public trust and allow for more rapid, but safer, innovation.

Conclusion:
Despite the fact that they’re created by humans, Large Language Models are still quite mysterious. This is why AI is often referred to as a ‘black box,’ a phenomenon that is difficult to understand from the outside, is the general perception. Recent research published by Anthropic, a leading company in the AI industry, aims to illuminate the more perplexing aspects of AI’s algorithmic behavior.

--

--

Pradeep Pujari

AI Researcher, Author, Founder of TensorHealth-NewsLetter, ex-Meta