Eric served the pot roast and watched Moretti take a careful bite.
“You know,” Eric said, “I’ve been practicing distant reading for about ten years now. Your idea. I took it to heart.”
Moretti set down his fork. “My idea was never about not reading. It was about reading differently. You can close-read one novel, or you can distance-read five hundred. The point was never to replace close reading — it was to make the scale visible.”
“Right,” Eric said. “And that’s exactly what the Distant Reader does. It takes hundreds of texts and models them. Word frequencies, topic models, network graphs — all the things you can’t see when you’re reading one book at a time.”
“But you still need the books,” Moretti said. “That was always the bottleneck. You can talk about distant reading all you like, but where do you get five hundred novels? You go to a library, and the library says: click, save, click, save.”
Eric laughed. “Exactly. That’s the problem I’ve been trying to solve. I built a collection of 1.7 million open-access texts so I’d have something to distant-read. But getting texts — that’s the part nobody talks about. The reading is the glamorous part. The acquiring is the soul-crushing part.”
Moretti nodded slowly. “In the Literary Lab, we had graduate students downloading things by hand. It was absurd. We were doing computational literary analysis and the bottleneck was a human being clicking ‘save’ four hundred times.”
“And authenticating,” Eric added. “Don’t forget authenticating.”
“How could I forget,” Moretti said. “I once had to authenticate through three different proxy servers to get a single nineteenth-century novel. Three. For a book published in 1847.”
Stallman, who had been eating in focused silence, looked up. “Authentication is a form of control. It has nothing to do with access and everything to do with power. The publishers want to own the text. The universities want to own the access. The software companies want to own the infrastructure. Everyone wants to own something. The only thing nobody wants to own is the problem.”
Hart chuckled. “That’s why I started Project Gutenberg. I didn’t want to own anything. I wanted to give away everything. The first document I typed in was the Declaration of Independence. I typed it in by hand, character by character, and I put it on the network for anyone to download. No authentication. No click, save. Just the text.”
“By hand?” Sam said. He looked stunned. “You typed it in by hand?”
“I typed it in by hand,” Hart said. “In 1971. On a Xerox Sigma V mainframe at the University of Illinois. People thought I was eccentric. They were right.”
Eric poured more wine and turned to Stallman. “Richard, I should tell you — in the essay I call it ‘open-source software.’ I know you prefer ‘free software.’”
“I do,” Stallman said. “Open source is a development methodology. Free software is an ethical stance. They are not the same thing. You can have open source without freedom. You cannot have freedom without the source.”
“I understand the distinction,” Eric said. “But I’m writing for librarians, not programmers. If I say ‘free software,’ they think I mean no cost. If I say ‘open source,’ they at least understand it’s about the code being visible.”
“Then educate them,” Stallman said. “Don’t dilute the message for convenience. The word ‘free’ has been misunderstood since 1776. That didn’t stop the founders.”
Hart smiled. “Richard, I once had someone ask me if Project Gutenberg was ‘free as in beer’ or ‘free as in speech.’ I said it was free as in library.”
Stallman considered this. “Free as in library. I haven’t heard that one.”
“It means: you don’t have to ask permission, you don’t have to pay, and you don’t have to explain yourself. You just walk in and read. That’s what a library is. That’s what I wanted Project Gutenberg to be.”
Eric said, “And that’s what my stacks are. 1.7 million items, no authentication, no middleware, just pure HTTP. You can crawl the whole thing with a spider if you want. No click, save.”
“How many of those 1.7 million came from Project Gutenberg?” Moretti asked.
“About 60,000,” Eric said. “Plus HathiTrust, EarlyPrint, Notre Dame pamphlets, COVID articles, and a lot of web-harvested scholarly papers. I’ve been collecting for thirty years.”
“Thirty years,” Hart said, shaking his head. “I collected for forty. We should have compared notes sooner.”
“We’re comparing notes now,” Eric said.
Eric turned to Kirsten, who had been eating quietly and methodically, the way someone eats when they are used to being overlooked at dinner parties.
“Kirsten, Index Data built the underlying indexer and SRU interface for Koha. Without your software, none of this works. You created the foundation.”
Kirsten set down her fork. “Index Data has been building open-source library infrastructure since the 1990s. Zebra, our indexer, was designed to handle MARC records and bibliographic data. YAZ, our toolkit, implements Z39.50 and SRU. We built these because we believed — and still believe — that library systems should be built on open standards, not proprietary lock-in.”
“And you’ve been doing this since before the phrases were coined,” Eric said, with a glance at Stallman.
Stallman raised his glass to Kirsten. “Index Data gets it. They built free software for libraries before it was fashionable. That matters.”
Kirsten nodded. “It matters because infrastructure is invisible. Nobody thinks about the indexer when the search works. They only think about it when it breaks. We built something that doesn’t break.”
“Koha just goes,” Eric said. “I installed it two years ago on a $25/month cloud server. I haven’t touched it since. Two cores, 4 GB of RAM, 80 GB disk. It indexes 1.7 million MARC records and it just goes.”
“That’s because Zebra is fast,” Kirsten said. “And because the MARC standard, despite being from 1965, is remarkably well-structured for indexing. People complain about MARC, but it works.”
“Like a free kitten,” Eric said.
“I beg your pardon?” Kirsten said.
“Open-source software is only as free as a free kitten. The software costs nothing, but the maintenance — feeding the kitten, cleaning up after the kitten — that’s not free. I’m paraphrasing myself.”
“That’s actually from me,” someone said.
Everyone turned. Andreas, who had been quiet until now, spoke with the careful tone of a man who is more comfortable writing patches than making conversation.
“The kitten metaphor. I read it on your blog. I liked it.”
“Well,” Eric said, “Andreas — you’re the reason my SRU server supports facets. Koha didn’t support faceted search results out of the box. I complained on a mailing list, and you wrote a patch. Just like that. You wrote it and it worked.”
Andreas shrugged. “The mailing list post was clear. You described the problem precisely. I had the context to solve it. It took an evening.”
“An evening,” Moretti repeated. He looked genuinely impressed, which for Moretti required a significant reconfiguration of his facial muscles.
“The patch is on GitHub,” Andreas said. “Anyone can use it. That’s the point.”
Eric turned to Sam. “And Sam — you wrote the MCP server for SRU. The whole thing. I found it, forked it, and adapted it.”
Sam shifted in his chair. “I wrote it because I thought SRU was underused. It’s a good protocol, but it’s hard for non-programmers. The CQL syntax alone is enough to scare most people away. I thought: what if an LLM could handle the syntax for you?”
“And that’s exactly what it does,” Eric said. “I say ‘Find articles about love,’ and your MCP server translates that into CQL, sends the query, gets the XML, and the LLM reformats it into plain English. No Boolean logic. No fielded searching. No CQL. Just a question and an answer.”
“But you forked my code,” Sam said, with a small smile. “You changed it.”
“Your server didn’t support facets,” Eric said. “And it included functionality my SRU server didn’t support, like the scan operation. And it imposed default values that were invalid for my implementation. So I forked it. That’s how open source works.”
Sam nodded. “That is how open source works. You fork, you adapt, you contribute back. I’m glad you found it useful.”
“I invited you to dinner, didn’t I?”
“You did.”
Moretti poured himself a second glass of wine and leaned back. “Let me make sure I understand the whole architecture. Eric has 1.7 million open-access texts. He cataloged them with MARC records. He indexed them with Koha, which uses Index Data’s Zebra indexer. He exposed the index through SRU, which returns XML instead of HTML. Andreas patched SRU to support facets. Then Sam wrote an MCP server that wraps the SRU API, and an LLM sits between the person and the server, translating natural language into API calls and interpreting the results.”
“That’s it,” Eric said.
“And the LLM evaluates the search results,” Moretti continued. “Not just retrieves them — evaluates them. It summarizes, classifies, identifies themes, generates bibliographies.”
“Yes.”
“This is not distant reading,” Moretti said. “This is something else. Distant reading is quantitative. You count things. You graph things. The patterns emerge from the data. What you’re describing is interpretive. The LLM is not counting — it’s reading and summarizing. That’s a different operation.”
Eric considered this. “You’re right. It’s not distant reading in your sense. It’s more like… a reference librarian. A reference librarian who has read everything and can tell you what they found.”
“A reference librarian who hallucinates occasionally,” Stallman said.
Everyone looked at him.
“You said the output contains ‘very, very few inaccuracies,’” Stallman continued. “Very few is not zero. In a bibliographic context, one inaccuracy can send a researcher down the wrong path for months. What guarantees do you have?”
Eric was quiet for a moment. “None. That’s honest. The LLM is non-deterministic. I can run the same query twice and get different summaries. And I can’t promise it won’t fabricate a facet value or misattribute an author.”
“So how do you trust it?” Stallman asked.
“I don’t trust it completely,” Eric said. “I trust it the way I trust a first draft. It gives me a starting point. It gives me an overview. It saves me time. But I verify. The CSV file it generates — I curate that. I remove duplicates, normalize names, check dates. The LLM does the first pass. I do the second pass.”
“That’s a reasonable answer,” Stallman said, though his expression suggested he found it only partially satisfactory. “But you’re building on Anthropic’s protocol. MCP is proprietary. Claude is proprietary. The LLM is a black box. You’re advocating open-source principles and then building your flagship tool on closed infrastructure.”
Eric nodded. “That’s a fair criticism. And I don’t have a good answer. MCP is new, it’s from Anthropic, and if they change it or deprecate it, I’ll have to adapt. But the alternative — waiting for an open-source LLM protocol — could take years. I’d rather build something now, demonstrate it works, and let the open-source community create alternatives.”
“Build now, open later,” Hart said. “I like that. That’s what I did with Project Gutenberg. I typed the text in by hand because there was no alternative. If I’d waited for an open-source text digitization standard, I’d still be waiting.”
“The point is to show it works,” Eric said. “To show that a library catalog can do more than return results. It can evaluate results. It can summarize. It can classify. It can generate a bibliography in Chicago style. And it can do all of this from a natural language command.”
Moretti swirled his wine. “And you believe this will come to traditional library catalogs.”
“I do. It’s already happening on the consumer web. Google does AI overviews. DuckDuckGo does AI summaries. The library world is behind, but it won’t stay behind forever. The question is whether libraries build this functionality themselves — on open-source foundations, the way Index Data built Zebra and Koha built the ILS — or whether they license it from vendors.”
“And you hope they build it,” Kirsten said.
“I hope they build it. I hope they build it collaboratively, on open standards, before they feel forced to buy it. Because once they buy it, they lose control. And libraries losing control of their own infrastructure — that’s a tragedy I’ve seen before.”
Stallman raised his glass. “On that, at least, we agree.”
Hart broke the silence. “Eric, you said the collection is designed to stand the test of time. ‘As long as there are computers connected via TCP/IP and those computers support HTTP.’ You said you wonder how long that will be.”
“I did.”
“And you said the Internet is fragile.”
“It is.”
Hart tore a piece of bread. “I used to think about this. I still think about it. Project Gutenberg has survived because the texts are plain. ASCII. UTF-8 now. No DRM, no proprietary format, no authentication. You can put ten thousand books on a USB drive and walk away. If the Internet dies, the books survive.”
“Your stacks are the same,” Eric said. “Pure HTTP. No middleware. No PHP, no Python scripts to break. Just files on a server.”
“But the server is on the Internet,” Hart said. “And the Internet is fragile.”
“Probably not as long as codices,” Eric said. “Books have lasted centuries.”
“Codices,” Moretti said, smiling. “You know, I once held a first edition of I Promessi Sposi in a library in Milan. The binding was falling apart. The pages were yellow. And it was still readable. Five hundred years old and still readable. Will your server be readable in five hundred years?”
“Almost certainly not,” Eric said.
“Then print it out,” Hart said. “Print the important ones. Keep them on a shelf. The cloud is not permanent. The shelf is permanent.”
Eric laughed. “Michael, you are the only person I know who would tell me to print out 1.7 million documents.”
“I’m not telling you to print 1.7 million. I’m telling you to print the important ones. The ones that answer your big questions. Truth, beauty, honor, justice. Print those. Put them in a box. Give the box to a library.”
Stallman, who had been listening with unusual patience, said: “The question of permanence is a question of freedom. If the format is free, the text can be migrated. If the format is proprietary, the text is trapped. Plain text is the freest format. It has outlived every proprietary format ever invented. It will outlive HTML, XML, JSON, and whatever comes next.”
“So the answer is plain text,” Moretti said.
“The answer is always plain text,” Hart said.
Eric served dessert — a simple apple tart — and the conversation turned forward.
“What’s next?” Moretti asked.
“I need to integrate the bulk download script into the MCP server,” Eric said. “Right now, I can search and evaluate, but I can’t acquire in bulk through natural language. Once I can say ‘Cache the items found from this search,’ the loop closes. No more click, save.”
“And then?” Kirsten asked.
“And then I hope someone — some enterprising library student, or some library, or some library consortium — takes the idea and applies it to their own catalog. Wraps their SRU or Z39.50 endpoint in an MCP server. Lets people search with natural language. Lets an LLM evaluate the results. Builds it on open-source foundations.”
“You’re issuing a challenge,” Moretti said.
“I’m issuing a challenge and a hope. The challenge is: build this. The hope is: build it openly.”
Andreas, who had been quiet again, said: “The patch I wrote — I wrote it because someone described a problem on a mailing list. That’s how it works. Someone describes the problem. Someone else solves it. You don’t need permission. You don’t need a committee. You just need a mailing list and an evening.”
Sam nodded. “And a fork. Sometimes you need a fork.”
Stallman, who had been frowning thoughtfully, said: “The thing I want to know is: when the LLM summarizes the search results, where does the summary go? Does the person keep it? Does the server keep it? Is it logged? Is it private?”
Eric paused. “I… don’t know. The MCP session runs locally. The LLM output is saved as an HTML file on the person’s machine. But the LLM itself runs on Anthropic’s servers. The query passes through their API. So yes, Anthropic sees the query.”
“That’s a problem,” Stallman said.
“It’s a problem I acknowledge,” Eric said. “And it’s part of the tension I’m living with. I want open-source infrastructure, but I’m using a proprietary LLM. I want privacy, but the LLM is cloud-based. I want libraries to build their own, but the LLM technology is not something most libraries can host.”
“Yet,” Kirsten said.
“Yet,” Eric agreed. “The ‘yet’ is doing a lot of work in that sentence.”
Hart said: “I remember when the idea of a digital library was impossible. People said it couldn’t be done. People said the storage alone would cost millions. People said the texts would never be digitized. And yet. Here we are. 1.7 million texts on a $25 server.”
“Yet,” Eric said.
Eric looked around the table — at the Italian literary critic, the free-software crusader, the quiet visionary who had typed the Declaration of Independence by hand, the Greek developer who had spent an evening writing a patch, the young Australian who had written an MCP server because he thought a good protocol was underused, and the Danish infrastructure engineer who had spent fifteen years building things that don’t break.
“I’ve been trying to get us all in the same room for about a decade,” Eric said. “Some of you I’ve never met. Some of you I can’t meet. And yet here you are.”
“You already said that,” Moretti said. “At the beginning.”
“I know. I’m a librarian. I repeat things for emphasis.”
Hart smiled. “Eric, I smelled the bread from the driveway. That was enough for me.”
Stallman raised his last glass. “To plain text. To free software. To infrastructure that doesn’t break.”
Kirsten raised hers. “To open standards.”
Andreas raised his. “To mailing lists and evenings.”
Sam raised his. “To forks.”
Moretti raised his. “To distant reading — and whatever comes after it.”
Hart raised his. “To giving things away.”
Eric raised his. “To all of you. To dinner.”
They drank.
Afterward, Eric cleared the table, washed the dishes by hand, and sat down to write. The bread had been good. The conversation had been better. The software, as always, was not yet done.