In a truly inspiring and innovative development, software developer Ishan Anand, a self-declared Excel enthusiast and spreadsheet whiz, has managed to integrate the GPT-2 language model into Microsoft Excel. This groundbreaking accomplishment not only showcases the remarkable versatility of spreadsheets but also offers an intriguing peek into the inner workings of large language models (LLMs) and their underlying Transformer architecture that powers intelligent next-token prediction.
Anand’s Revolutionary Approach
Recognizing the profound intricacy of ai systems, Anand asserts that mastering a spreadsheet is the key to unlocking the mysteries of artificial intelligence. With confidence in his voice, he declares, “If you can understand a spreadsheet, then you can understand ai.” This visionary approach has led to the creation of an impressive 1.25GB spreadsheet, which Anand generously shares on GitHub for anyone interested in downloading and exploring further.
Although Anand’s GPT-2 implementation in Excel may not rival the advanced capabilities of modern LLMs, it offers a unique perspective on this groundbreaking model. It is essential to remember that GPT-2 was introduced in 2019 and predates the conversational ai era, with ChatGPT emerging from attempts to elicit conversational responses using GPT-3 in 2022.
Exploring the Transformer Architecture
At the heart of Anand’s Excel integration lies the GPT-2 Small model, a 124 million parameter strong contender. Compared to the full version of GPT-2 with its impressive 1.5 billion parameters and GPT-3’s even more astounding up to 175 billion, Anand’s implementation still manages to illustrate the transformative power of the Transformer architecture in performing intelligent “next-token prediction.”
Though this spreadsheet can handle a mere 10 token input, a trivial fraction compared to the impressive capacity of GPT-4 Turbo’s 128,000 tokens, Anand’s work offers invaluable educational value. He believes his “low-code introduction” is the perfect resource for tech executives, marketers, product managers, ai policymakers, ethicists, developers, and scientists seeking to deepen their understanding of LLMs’ foundational principles.
A Foundation for Modern LLMs
Anand maintains that the Transformer architecture employed in his GPT-2 implementation remains “the foundation for OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Bard/Gemini, Meta’s Llama, and many other LLMs.” His multi-sheet project provides users with a step-by-step guide through word tokenization, text positions and weightings, iterative refinement of next-word prediction, and ultimately, selecting the output token – the predicted last word in the sequence.
One significant advantage of Anand’s Excel-based implementation is its ability to run the LLM entirely locally on a PC, without requiring cloud services or API calls. However, Anand advises against attempting to use this Excel file on Mac or cloud-based spreadsheet applications, as it might result in crashes and poor performance. Furthermore, he strongly recommends using the latest version of Excel for optimal performance.
Though Anand’s GPT-2 implementation might not reach the capabilities of contemporary LLMs, it stands as a remarkable educational tool and a testament to the unyielding adaptability of spreadsheets. By shedding light on the inner workings of language models, Anand’s work empowers individuals from various backgrounds to gain a more profound understanding of artificial intelligence and its underlying structural principles.
A Lasting Legacy
Anand’s integration of the GPT-2 language model into Microsoft Excel sets a precedent for merging cutting-edge ai technology with ubiquitous spreadsheet software, paving the way for future innovations and demystifying the complexities of LLMs. This groundbreaking achievement offers an engaging, educational experience that appeals to both seasoned professionals and curious beginners seeking to explore the vast potential of artificial intelligence.
