2017: Attention
Attention Is All You Need replaced recurrence-heavy sequence modeling with self-attention, where tokens compare against other tokens and mix information through learned matrices.
Spreadsheet model release
SAAYN GPT-2
A local Python wrapper around spreadsheets-are-all-you-need, packaged with OpenAI-style API endpoints, a persistent CLI chat script, streamed token visualization, and optimized headless execution.
Published artifacts
4 zips
Largest portable bundle
2.5G
Faithful source
Workbook
Fast path
CSV + NumPy
Working Browser Lab
Talk to the spreadsheet-derived model
This form calls the live same-origin SAAYN API at /api/saayn/v1/chat/completions.
The backend is the Python wrapper around the spreadsheet GPT-2 weights, using the optimized NumPy path rather than a visible spreadsheet window.
Checking API...
engine: unknown
model: saayn-gpt2-spreadsheet
assistant
Type a prompt below. Chat mode adds a small dialogue memory primer; completion mode sends only your text as a raw GPT-2 continuation prompt.
Why this artifact matters
From attention to a haunted accounting ledger
SAAYN sits on a historical loop: the Transformer made language modeling matrix-native, GPT-2 made raw next-token prediction culturally visible, the spreadsheet implementation unfolded that machinery into cells, and this wrapper turns the educational artifact back into a usable local runtime.
2019: GPT-2
GPT-2 used a decoder-only Transformer as a continuation machine: given prior tokens, predict the next token. OpenAI initially staged its release, then published the full 1.5B model in November 2019.
Spreadsheet fossil
Spreadsheets Are All You Need makes the hidden matrix math legible by translating token embeddings, attention, MLP layers, and logits into spreadsheet grids and formulas.
How the spreadsheet worked
The transformer path, cell by cell
The original workbook is educational because it refuses to hide the pipeline. A token becomes an integer id, that id indexes an embedding row, positional information is added, and then each Transformer block repeats attention plus an MLP until the final logits choose the next token.
Tokenization
Text is split into byte-pair tokens. In this live wrapper, each streamed row shows the chosen token and its vocabulary id.
text -> BPE tokens -> token ids
Self-attention
Each token creates query, key, and value vectors. Query-key similarity decides what previous tokens matter for the next prediction.
Attention(Q,K,V) = softmax(QK^T / sqrt(d_k))V
Logits
The final hidden vector is projected back into vocabulary space. The wrapper uses greedy argmax, so the largest logit becomes the next token.
hidden -> vocabulary logits -> argmax token
The public page is not running Excel in your browser. It talks to a headless Python process that directly uses the spreadsheet-exported GPT-2 parameters. The workbook remains included in the full/runtime bundles for provenance; the fast path uses the CSV parameters and NumPy cache.
Bundle Matrix
The workbook is preserved where it matters, but the fastest runtime path uses the extracted GPT-2 CSV parameters and NumPy cache without opening a spreadsheet window.
| Bundle Variant | Size | Workbook | CSV Params | .venv | Offline Runnable |
|---|---|---|---|---|---|
| Ultra-light source | 2.3M | No | No | No | No. Downloads dependencies and GPT-2 CSV params during install. |
| Lite | 577M | No | Yes | No | No. Model artifacts are included, dependencies still install locally. |
| Full | 1.9G | Yes | Yes | No | No. Includes workbook and params, not a complete runtime. |
| Runtime Linux | 2.5G | Yes | Yes | Yes | Yes, on similar Linux/Python 3.10 x86_64 systems. |
Ultra-light Source
2.3MSource only. The install script downloads Python dependencies and GPT-2 CSV parameters.
Downloade516283869dcef326c23855b26c63528c03988b27c7587c96b807b2b44419927
Lite
577MIncludes CSV model parameters for the optimized headless path.
Downloada28f9892882cf2f22b1e695c01bf5a1c8b8720192a26fc1f02482de1e704afc9
Full
1.9GIncludes the original spreadsheet workbook plus extracted CSV parameters.
Download52159a2a161a819b559f1dbc226258f2d0f127c28662b1aa2e65f2d35ed4384f
Runtime Linux
2.5GOffline-oriented Linux runtime with workbook, CSV params, and a bundled virtual environment.
Downloadf62f93dc85dbf6dde744b10f0c64822162f56de42ac2262568f54dbf57ecb7ea
What it runs
OpenAI-style local interface
The Python wrapper exposes modern OpenAI-compatible completion and chat-completion routes, plus a CLI chat loop with persistent session memory.
Chat mode defaults to assistant/user dialogue prompting. Plain current-input completion is available with --normal-completion, /completion on, or /mode completion.
How it stays faithful
Spreadsheet behind the scenes
The original workbook is included in the full/runtime archives for provenance. The optimized engine directly uses the same GPT-2 weights exported from the spreadsheet representation, avoiding an interactive spreadsheet window while preserving the spreadsheet implementation lineage.
Smoke check: prompt Mike is quick. He moves should emit quickly as the next token on the optimized engine.
Sources and references
Primary lineage links
Transformer paper
Attention Is All You Need describes scaled dot-product attention and the Transformer architecture.
GPT-2 release history
OpenAI's GPT-2 announcement and 1.5B release note document the staged release.
Spreadsheet implementation
ianand/spreadsheets-are-all-you-need is the source lineage for the workbook artifact wrapped here.