Ben Brumfield

We'll try to write up some documentation once the feature settles down; there are still some more changes we want to make, possibly before tomorrow.

(The webinar will be recorded and sent to anyone who signs up, though.)

Introducing Gemini 3.0 Support in FromThePage - FromThePage Blog When Ben sent me Mark Humphries’ report on testing a new, unreleased Gemini model, I got scared. And excited. Mark is a historian and digital humanist who’s gone deep on analyzing AI tools for textual...

We wrote this two weeks ago, but it's sadly outdated since it doesn't include any of the accuracy statistics, the reasoning display, or the tuning we've done to the prompt since then:

content.fromthepage.com/introducing-...

I'll try to find a link to a live page.

"From a transcribers point of view I will be very unlikely now to continue devoting time to working on straightforward handwritten documents without an AI draft as a starting point ."

She plans to attend our webinar this Thursday and might be willing to talk about her experience.

An update: after a week of testing, the same volunteer is now very happy:
"I have been very pleased to use the AI facility in transcribing Nicholas Piper Log books for Whitby Literary & Philosophical Society."
...

Webinar link: content.fromthepage.com/dec-2025-web...

"From a transcribers point of view I will be very unlikely now to continue devoting time to working on straightforward handwritten documents without an AI draft as a starting point ."

She plans to attend our webinar this Thursday and might be willing to talk about her experience.

An update: after a week of testing, the same volunteer is now very happy:
"I have been very pleased to use the AI facility in transcribing Nicholas Piper Log books for Whitby Literary & Philosophical Society."
...

Screenshot showing "3,008,092 Transcribed Pages"

It looks like we passed another milestone on FromThePage last week:

But it really feels like libraries & archives as a field suddenly just went from "we aren't generally attempting to do automated handwriting recognition because it's at the edge of what's possible" to "oh boy now we have another doable but labor-intensive collections enhancement task on the backlog"

I was doing some experiments this morning as we fine-tune our prompts and am astonished to say that while I've seen many errors produced by Gemini, none of them were the seductively plausible hallucinations that have made me regard MMLLMs as potentially poisonous.

We're working on a pipeline for extracting structured data from industrial drawings into structured data fields. More in January, but so far we're very impressed with LLMs for this purpose.

Wow.

People who like to transcribe tend to be hands-on types, or puzzle solvers, or people who read between the lines. Transcribing is a way of thinking. It is not for every project, but for some projects, it can be crucial.

As someone with FtP projects, I think there are still several good reasons for people to choose transcribing. The goal is not necessarily to record the words, but to find meaning. Sometimes that comes from reading the words, other times it comes from closely reading each mark.+

I have never felt that AI is a threat to transcription projects. Transcription is such a fulfilling experience and the words run through you in ways that reading alone can never do. Meeting up with scribes of the past will always be thrilling.

the projects that are successful.

That said, I don't relish saying, "Stop doing [thing you enjoy] and instead do [totally different thing]."

It's not yet clear that we'll need to do so, since we still need to do our own evaluation of Gemini's capabilities and weaknesses. Our hope is that it will open up collections that were unsuitable for crowdsourcing (or at institutions with no public mandate/volunteer expertise) without replacing(+)

There may be some non-engagement follow-on benefits as well. I have had two volunteers write that transcribing helps their anxiety, which I can certainly see.

"I have very much enjoyed the opportunity of working on historical material over the past 5 years but today feel dismayed that I may now be wasting my time in continuing."

Sure enough, we got this email from a volunteer on Friday after we announced Gemini integration:

"As a long-term transcriber on From the page I am now wondering about the implications of Gemini 3 (AI) for me - I am feeling particularly discouraged today.

Last week, we introduced support for #Gemini3 into our collaborative #transcription platform FromThePage. We are still experimenting with prompts and outputs, but we think that this may be the first #LLM we've seen that does not introduce seductively plausible errors into historic documents.

Introducing Gemini 3.0 Support in FromThePage - FromThePage Blog When Ben sent me Mark Humphries’ report on testing a new, unreleased Gemini model, I got scared. And excited. Mark is a historian and digital humanist who’s gone deep on analyzing AI tools for textual...

Introducing #Gemini 3 Support in #FromThePage content.fromthepage.com/introducing-...

We're still developing capabilities and guardrails now, but plan to present it all at a webinar December 11.

One library and one scholarly edition are testing it out so far, and we hope to learn from them as well as from our own experiments.

We'll be writing more about our experiments and findings about accuracy, plausibility, and user experience over the next week or so.

Yesterday, Google released Gemini 3, which has gotten really interesting reviews from Mark Humphries: generativehistory.substack.com/p/the-sugar-...

Also yesterday, we shipped an integration between FromThePage and Gemini, allowing transcribers the option of starting with an AI draft.

Is That Transcription Really Human? - FromThePage Blog Last month, someone asked this question on the Genealogy and AI Facebook group:If volunteers use AI to transcribe documents, is that OK? I have strong opinions, but want to explain them. First off, th...

Should a volunteer use #AI to help them transcribe pages for a #crowdsourcing project? That question got me thinking about why, exactly, my answer is "no" and what kinds of purposes different transcriptions may be used for.

content.fromthepage.com/can-voluntee...

The Legacy of India’s Quest to Sterilize Millions of Men Breaking the Cycle: Part 1 In 1976, men across India were drastically changing their behavior. Some were abandoning the beds inside their homes to sleep in fields; others were skipping major festivals...

This one made a big impression on me: pulitzercenter.org/stories/lega...

However this might be a bit closer to the time-frame you're looking for: www.theatlantic.com/magazine/arc...

I'd love to hear how and whether other developers in #digitalhumanities or libraries and archives have done similar experiments and evaluations.

For now, we're exercising a lot more discipline with these tools to keep from wasting our time on shiny new things.

3. Pitching in on specific tasks that we don't have enough skills to do as well ourselves, like translating messages for our recurring "Werk/Arbeit;obra/trabajo" problem or tweaking our UI for A11Y issues (which it does well with).

2. Helping refactor or isolate our legacy test suite. Asking an agent to isolate a single test at a time might finally get us out of dependency hell in our test suite.

1. Fixing small user-reported/developer-noticed problems so that we don't have to interrupt developers in other effort. This lets us fix bugs on a managers schedule instead of a maker's schedule (cf. Paul Graham) . (Sara and I spend most of our days on a managers schedule, unfortunately)