Large Learning Models

“What’s at risk is nothing less than the ability to think for oneself. If people become overreliant on AI to write for them, they could lose the capacity to interrogate or even develop their own views. These are quintessentially human capacities.” Rose Horowitz, ‘The end of reading’, The Atlantic

A bit of background. I trained and started my career as a teacher. I am now lucky enough at the Guardian to work with a range of brilliant product managers, developers, data scientists and designers. I am accustomed to the language of product building and the processes around it. But I can't code myself. Over the past three years, my work has been focused on using LLMs in different contexts to understand their limitations and the risks associated with them. These are all very real. Any engagement with LLMs requires care.

But too often the conversation around generative AI only focuses on risk. And in doing that we end up considering negative change with none of the potential upside. It's also widely assumed that personal and professional responsibility is somehow automatically bypassed when it comes in contact with these machines, rather than recognising the fact that any professional values, codes and good practice can and must be applied in exactly the same way to LLMs as any other input or tool you might use.

This piece isn't for people already building with LLMs. I'm far from the bleeding edge. But I wanted to write it to demonstrate just how rich, educative, provocative, thoughtful and valuable building with an LLM can be, rooted in a practical example that I started working on just a week ago.

The starting point

My daughter recently studied Dracula at her secondary school and is now reading David Copperfield. All of this started from a conversation at breakfast about how hard it can be to get a concrete sense of the world around these huge novels, and led to me suggesting that we try to trace some locations and journeys on a modern map. We started with Bram Stoker.

I happen to use the Claude desktop app to work with Claude Code to build things. If you're unfamiliar with all this, it simply means that my interaction with Claude is through talking to it - just like a typical LLM. Claude Code is a 'harness' that lets Claude create code either on your computer or in a public digital space. Many people assume that to make something impressive requires hours of training or watching YouTube tutorials. These things can help. But you can also just ask the model itself. Most of my projects start with a question along these lines:

"I'm thinking of trying to build a map of Dracula's locations and some of the journeys of the characters. I'd really love to use a contemporary map. Is that possible? What things do I need to consider? What's the easiest way to do it? I can't code. You will need to be super clear and patient."

The first lessons

I'm not going to go into the technical details of this project. We learned a lot about the technical aspects of engineering something like this as we went, but what mattered more was what we learned about the topic we were exploring. Conversations about LLMs in education, especially between those that don't use them much, tend to start and end with the idea of cheating or avoiding thinking by shortcutting. I hope this example might underline why there is far more value to discover.

Right off the bat, Claude started to highlight a bunch of things we needed to consider and make calls on to shape the finished product. It found a way to lay the actual Ordnance Survey of the 1890s - scanned and georeferenced by the National Library of Scotland - over a modern map, so that across Britain Dracula plays out on the survey of its own decade. It explained where that map came from, and how we were allowed to use it. It talked about why it existed (a lovely detour into maps, the British empire, colonialism and Brian Friel's Translations followed in person) and how we were able to use it. It also drew out the distinction between real locations and imagined ones. Tate Hill Pier is a real place, identifiable to the metre. Castle Dracula is somewhere "above the Borgo Pass". Claude suggested taking best guesses, building on detail from the novel, and distinguishing between these things on the map.

Tracking the journeys of the characters felt like a simple suggestion but was in fact a deep, deep well of decisions and considerations that ended up being the engine of the whole idea. I told Claude that I was imagining "Indiana Jones crossed with Bram Stoker". People are quick to dismiss LLMs' writing abilities, but a long and detailed conversation with one will generally leave you reeling from just how nuanced and flexible these tools can be in inferring intent and responding. That simple expression of the idea requires a huge amount of decoding. Claude could immediately contextualise it and extrapolate out.

Building it required addressing a series of big questions:

Choosing who the characters are that drive the narrative, especially through movement;
Identifying all the key journeys and creating a script that gave context to each movement. Claude suggested double checking every quote against the original text;
Thinking about how to handle timelines, chapters and flashbacks.

All this eventually got us to a simple interactive animated map that did a nice job of showing where characters travelled, went some way to articulating how and why, and started to illuminate the chronology of the book. Crucially, while Claude was building this for us, our involvement was not passive or mindless. The process and the decisions we had to make made this a lesson in itself: rich, deep and thoughtful engagement with the novel.

And this practical, iterative approach of starting small, watching the output and adjusting meant that we could shape the user experience as we went. My Indiana Jones prompt was a great way to start, but it also showed that tightly following a single character meant that we lost the value of the context of the wider map. Harker's travels from Munich to Castle Dracula made no sense zoomed in. We needed a "director" that took the script and the journeys and thought about how to frame each moment to deliver the geographical context.

The second book

It felt like what we had built could be even more interesting if the tool could show a user more novels and journeys. So we went after a novel that looked very different to Stoker's. Tess of the d'Urbervilles is rural, tightly focused and, of course, operates in a very different geographical reality: Hardy's Wessex is a map superimposed over a real one.

Claude immediately identified the challenge and the need to look at Hardy's own maps alongside a century of academic work pinning down the real locations of his world. There remained two genuinely unsettled places; Claude made a best guess and flagged them for what they were.

This is also where a serious conversation about copyright was initiated. I'd stated at the outset that we wanted to stay well within a safe area here. Claude, intuiting that this project was going to become something bigger, recommended harder and faster rules. We learned about UK copyright vs US copyright and elected to go with the former approach for everything that followed.

The third book… and four agents

With decent results on Dracula and Tess, we wanted a new challenge. Pride and Prejudice was a heart pick, but one with some really interesting and different challenges for the system we were building. It's set in the real world, and has some specific real locations, but locating Longbourn and other key places would rely on academic research and Austen's scrupulously stipulated distances.

One particular journey would make a huge difference to where the project went next. Lizzie's curtailed trip to the Lake District with her aunt and uncle leapt out on our map because Austen names the towns they travel through on their way. It bends through real places and comes alive. But many of our journeys, where only the start and finish were identified, were straight lines. We asked Claude if it was possible to find information about period coaching routes, emigrant-voyage records and more to bring those lines alive in reasonable ways, while always giving precedent to the authors' specifications.

And this is where the project and its engine became agentic. There's a lot of talk about agents at the moment. Most of it frankly clouds the main issue. It's all very modish and too hypothetical. The easiest way to think of an agent is an LLM that's been given a clearly defined job with clearly defined rules. The three examples we'd worked on already meant that our conversation had become a rich set of questions, rules and ideas that we could easily formalise into processes that could govern the transformation of any novel. I didn't really ask Claude to build agents: I just asked it to capture these things as separate instructions that it could follow in the future and build on as we tackled new texts.

You can see how our agents break down in some detail here. But broadly we needed four:

Research & curate locations, characters and quotes;
Plan the routes according to the most relevant period information;
Create a script that captures journeys and crucial narrative beats outside them;
Find relevant period images that could illustrate a given location.

Every novel that followed was produced this way. Any important issues Claude raised pre-publication. We then checked them live. Inevitably in each case we saw an issue (presentational, textual, temporal or geographical) that the specific novel threw up. Each time we learned something and fed that learning back into the agentic system to help it cope the next time it recurred.

Interventions and ideas

I hope you can already see the ways that detailed conversations lead to both participants - human and machine - bringing more and more richness to the project. This isn't an obvious thing to most people. Generally discussions around LLMs focus on bias and risk. But to finish off, here are just a few moments when Claude initiated conversations or suggested ideas that might surprise you:

When we created a bookshelf interface for the user to choose a novel, Claude rendered Dracula with a yellow cover and red writing. I hated it and asked for less garish colours. Claude pointed out that it had chosen them specifically because that was the colour scheme of the first edition.
The more you work on a novelistic map for this period, the more the colonial background becomes clear. We wanted to try to build in books that went beyond Britain and Europe. But most of those carry challenging frames, details, languages and concepts. When we pushed for novels that moved into Africa, India and South America, Claude specifically called this out. It led to a sensitivity pass that didn't aim to sanitise characters or actions but did look for real-world harms - a slur, an imperial framing - that would otherwise slip into our narration unremarked.
Attempting to fill some of the gaps on our map also led to Claude suggesting works we weren't familiar with. Spotting a void in Northern Europe, it recommended The Wonderful Adventures of Nils, a geography primer in the form of a novel by Selma Lagerlöf, the first woman to receive the Nobel Prize in Literature. More wonderful suggestions would follow, none of which I had read, and all of which are now on my reading list: Natsume Sōseki's Botchan, Ameen Rihani's The Book of Khalid and the glorious gothic doppelgänger yarn The Devil's Elixir.
We held off on War and Peace for a long time, recognising the scale of the challenge. Claude's key intervention here was wonderful: treat the Grand Army as a single protagonist. And once it was plotted on the map, a further refinement: it recommended using something called a Minard line - where the width of the trail indicated the size of the army. That's a spectacular idea that communicates in some way the scale of the losses of Napoleon's grandest folly. And one that would never have occurred to me.
In Search of Lost Time felt like something we needed to consider at least. But Claude had a strong view on the matter. "This book's whole argument is that place is subordinate to time - a paving stone or a madeleine matters not for where it is but for the buried years it releases. A faithful map would be eight or ten pins and a couple of train journeys. A reader would learn almost nothing true about the book. Worse, it might mislead." I think Claude was right. So we've swerved Proust for now.

So what are we left with? I think PlotLines is a lovely tool in and of itself. The comparisons between different novels, the telling of the stories and the ability to just explore the map and the places are all joyful and educative. But the real learning is exactly where most people see a blank: in a conversation with a machine.

❧