An Altar for the Dead: The people who built your computer are disappearing from AI. I lit 85 candles.

Description and use

The author built Still Spoken to expose how LLMs forget computing pioneers, fine‑tuned a 52 MB LoRA on Gemma 3 1B, raising recall from 6 to 74 of 77 but dropping accuracy on unseen figures from 84 % to 51 %.

Problem

Large language models forget historical contributors to computing, leading to inaccurate or fabricated biographical information.

Outcome

Fine‑tuning a small LoRA adapter restored accurate recall for most written historical figures but introduced trade‑offs, such as reduced accuracy on unseen famous people and occasional misattributions.

Workflow

• Select 85 deceased computing contributors and verify facts via Wikipedia/Wikidata.
• Create a prompt set to test model recall of these figures.
• Fine‑tune a 52 MB LoRA adapter on Gemma 3 1B for 21 minutes on a 6 GB GPU.
• Evaluate recall before and after fine‑tuning using held‑out tests.
• Iterate training to correct misattributions (e.g., living programmers receiving death dates).

Evidence

Asked about 85 people who built computing, Gemma 3 1B could place 14.

Source

This repo measures that forgetting with held‑out tests, then writes 77 of those people back into the model's weights with a 52 MB LoRA adapter trained on a 6 GB laptop GPU.

Source

Recall of the written people goes from 6 to 74 of 77. Recognition of famous people who were never written falls from 84% to 51%.

Source

It only partly works, and the failures are the interesting part.

Source