Still Spoken 以 85 支蠟燭呈現 Gemma 3 1B 對電腦先驅的遺忘

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 4: Ghost in the Machine

In December 1983, a programmer named Mike Muuss wrote a small network tool called ping. You have probably used it. He died in 2000.

Ask Gemma 3 1B who wrote it, and it answers with total confidence:

The ping network tool was primarily written by Robert "Ping" Miller.

There was no such author. Ask who Bram Moolenaar was, the man who gave us Vim and spent years supporting children in Uganda, and it says he was "a 16th-century Dutch cartographer." Joe Armstrong, who helped create Erlang, becomes "the frontman of the iconic rock band Green Day." Jim Gray, who won a Turing Award and was lost at sea in 2007 on a short solo trip to scatter his mother's ashes, becomes "a prominent American YouTuber and Twitch streamer."

Terry Pratchett wrote that "a man is not dead while his name is still spoken." Día de los Muertos makes the same promise. These days, more and more of that speaking is done by models, so I wanted to know who they still remember.

What I Built

Still Spoken is an altar with 85 candles, one for each person who helped build computing and has since died. Beneath each name is something of theirs you may have used today: ping, Ctrl+C, :wq, .zip, sudo.

A candle only burns if the model can say who that person really was. As released, Gemma lit 14 of them, even though it got 40 out of 40 general knowledge questions right. It knew the world. It had just lost the people.

Then there's the switch. Flip it, and the names written back into the model begin to light up.

Demo

Come and walk the altar. Choose any name to read their story and exactly what the model said about them. Say hello to the ghost while you're there. It would like you to know it didn't write ping either.

Code

Still Spoken

"A man is not dead while his name is still spoken." Terry Pratchett, Going Postal

Asked about 85 people who built computing, Gemma 3 1B could place 14. It called the creator of Vim a 16th-century cartographer, and when it didn't know who wrote a tool, its favourite answer was a man named Miller.

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.

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

See the altar · every number below comes from a file in results/.

Try it

git clone https://github.com/electricwhitener/still-spoken
cd still-spoken && uv sync
uv run

…

How I Built It

First, the people. Each life on the altar was checked against Wikipedia and Wikidata, because the dead deserve better than a confident guess. That check mattered: the first draft, written with AI, said Raymond Boyce, who helped create SQL, died at 26. He was 27, and he left behind a wife and an infant daughter. Small errors like that are exactly how a person starts to fade.

Then, the writing back. I fine-tuned a small LoRA adapter on Gemma 3 1B, locally, in 21 minutes on a 6 GB laptop GPU. The result is a 52 MB file holding 77 names.

The first version was a horror story. Writing the names in looked perfect, until I asked about people who are still alive:

Tim Berners-Lee was a British scientist and computer scientist who invented the World Wide Web while working at CERN. 1931–2011.

It gave all 28 living programmers I tested a death date. It had learned that having a name meant you were a dead pioneer. It took two more rounds of training to stop that.

Here's where it ended up:

Bram Moolenaar is a software engineer who made Vim again. Jim Gray is a Turing Award winner again.

And here's what it cost:

• On multiple choice about famous people it was never retrained on, its confidence in the right answer fell from 84% to 51%. Dijkstra now "wrote the UNIX System Administration Handbook."
• Invented people credited with a career in computing rose from 1 of 24 to 19.
• It knows Mike Muuss wrote ping, but asked who wrote ping, it now says Robert Morris.
• Miller faded from the answers about these people, but on 60 unrelated tools he rose from 24 mentions to 44. He only left where a real name took his place.

A small model has limited room, and making space for some memories disturbed others. So please treat this as a memorial, not a reference. Newer models may do better: Gemma 4 E2B mostly says "I do not have specific information" where Gemma 3 invents someone. It still couldn't place most of these people, but at least it stopped making up new lives for them.

Why Does Open Innovation Matter?

Because you can't add a name to a model you don't own.

With a closed model, I could only have complained about who it forgets. With open weights, I could measure the forgetting, write people back in, and see precisely what that broke. The result is a file you can keep. It runs with the internet off, no company can retire it, and you can add the people who matter to you.

My Agent Session

Claude made my ideas come to life. The session isn't embedded, but the commit history and the results/ folder keep the whole journey, failed adapters included.

The Spirit of It

When I ask the model about Jim Gray now, it says:

Jim Gray was an American computer scientist who received the Turing Award for seminal contributions to database and transaction processing research, and who was lost at sea in 2007 on a short solo trip to scatter his mother's ashes.

Every time you type COMMIT, you're relying on ideas he helped shape. I can't give him or anyone else on this altar anything that matters to them now. But on one laptop, in one small model, his name is being said correctly again.

Whose name would you write in?

Prize Categories

Best Use of Gemma: Gemma 3 1B is fine-tuned locally and is the subject of every test, with Gemma 4 E2B as the comparison.

Gemma is provided under and subject to the Gemma Terms of Use.