Spreadsheets are all you need
Implementing a full GPT‑2 small language model inside an Excel file has reignited interest in spreadsheets as a powerful, visual programming environment rather than just a business tool. Commenters highlight how this 1.25 GB, inference‑only model makes transformers and LLM internals far more tangible for learners, while also surfacing practical issues like Excel’s size and locale limits, lack of web/LibreOffice support, and the tiny context window. The project prompts broader reflection on when spreadsheets are ideal for rapid prototyping and end‑user computation, and when teams must graduate to more maintainable, conventional software systems.
Project & implementation
- The sheet implements GPT‑2 “small” fully in Excel, mainly as a teaching tool to show how transformers work.
- File size is ~1.2–1.25 GB and currently handles only about 10 tokens with a very short context window.
- It’s inference-only and lacks instruction tuning / RLHF and other refinements that turn a base LLM into a modern chatbot.
- There’s a GitHub repo and plans for more videos, including embeddings; Python-in-Excel could help with things like PCA and SVD.
Spreadsheets as programming / apps
- Many commenters say spreadsheets are their default prototyping or even production tool: from loan tracking, real-estate calculators, and internal tools to entire companies and trading desks.
- Spreadsheets are described as a low-friction, visual functional programming environment and even a “grandfather” of dataframes.
- There’s recurring tension: spreadsheets are extremely flexible and fast to iterate, but can become unmaintainable “hairballs,” after which teams migrate to more rigid internal systems and lose flexibility.
Excel skills & best practices
- Several people admit weak Excel skills despite technical roles, especially in finance contexts, and ask how to improve.
- Suggestions: learn by building real tools, reverse-engineer good sheets, use docs and books, and lean on LLMs for step-by-step help.
- Concrete tips: avoid hardcoded numbers, use colored input cells, master F4 for absolute/relative references, separate data from presentation, use INDEX/MATCH instead of VLOOKUP, exploit array formulas, LET/LAMBDA, named ranges, and basic VBA/SQL.
Visual programming & research context
- A long subthread argues spreadsheets are a major, widely deployed visual programming language.
- Cited reasons: direct manipulation of visible data, 2D layout, implicit dependency graph, immediate feedback, and avoidance of explicit variables.
- Others push back that formulas are mostly hidden, so they don’t match the “boxes-and-arrows” mental model, but are answered with a broader definition of “visual” centered on interaction style rather than code display.
Tooling, scale, and limitations
- Excel’s locale issues surface: function names and CSV separators differ by language; portability is problematic.
- The sheet doesn’t work in LibreOffice or web Excel due to size limits; Google Sheets can’t fit the full model.
- Alternative spreadsheet-like backends (e.g., Arrow-based, Row Zero) and GPU/distributed execution are discussed as ways to scale the paradigm.