The affordance of fun

An affordance is a set of possibilities offered by an environment. Or think of it as the range of possible actions that a user can perceive.

The first time I consciously came across this term was in The Organized Mind by Daniel Levitin, where he introduces the concept of Gibsonian affordances, named after the American psychologist James J. Gibson, who coined the term.

And it is a very cool term: affordance.

I’ve been thinking about it because right now AI is offering a whole bunch of affordances that I could have scarcely imagined otherwise, let alone acted on. And perhaps one of the most important affordances AI has offered me — and it’s not just me, because I’ve heard multiple people, including Venkatesh Rao, say something similar — is fun.

I’m having more fun than ever playing around with these tools: building small things, running experiments, doing things I’ve always wanted to do but didn’t have the technical ability to do.

At one point, I might add — I’m borrowing this phrase from Jasmine Sun — a lot of my interests, a lot of the things I wanted to do, were software-shaped. AI, naturally, considering how good these coding tools have become, lends itself to software-shaped ideas and software-shaped fun.

But just because the things I’m doing, or the fun I’m having, are software-shaped, doesn’t mean I have my head buried inside screens. The fun is leaking out into the real world. The projects I’m doing are opening up conversations with people I’ve never met in my life.

And it’s been fun.

The latest fun project I’ve been working on is around data visualization.

I recently started two projects. One is This Indian Life, and the other is Hope and Despair. These projects were an outgrowth of Project Akshara, another digitization project I was working on, where I was using large language models to extract text from old books in the public domain.

I was amazed at their vision capabilities.

Also, in my random browsing on the interwebs, because I work in finance, I keep looking at a variety of datasets from different organizations, a lot of them government datasets. And I kept coming across amazing data that wasn’t really available in a clean, structured, usable format publicly.

Because I had some experience extracting structured text from scans of old books, I knew that extracting data from tables was pretty much a cakewalk for most large language models.

I’ve always loved data visualization, though I didn’t really have the skills to do it well. But because I work in finance, data is front and center. Often, a good visualization can teach you more than a blog post of thousands of words. This has been my own experience. A lot of my understanding of finance has been visual in ways I hadn’t fully appreciated until I started writing this.

I had been doing basic visualizations at work for a variety of financial data and for the various things I wrote. And then I was kind of surprised at how good a lot of these datasets on unusable, horrendously designed government websites were.

So I figured I’d experiment with trying to visualize this data.

But instead of just creating another website with a bunch of charts, which a lot of people would frankly find hard to understand, I wondered: can I leverage LLMs and their natural language abilities to pull data, clean it up, visualize it, create high-quality charts, and then tell a story?

Can I explain what a particular data point means, put it in context, and present it in a way that even people who otherwise don’t follow the news, or aren’t statisticians, can understand?

That’s how This Indian Life, and by extension, Hope and Despair, were born.

I was kind of shocked at how easily I could do this because of the unique affordances offered by LLMs. They have access to tools. They can ingest large amounts of data, use data analysis libraries, mostly based on Python, and then leverage different visualization libraries to create visuals. And of course, they are really good at explaining things in any manner you want.

Well, in a loose sense. How well this works depends on your starting benchmark.

If you’re a normie like me, someone who doesn’t have programming skills, who is not a statistician, who is not really well-versed in statistics apart from basic arithmetic, and who isn’t a historian, so to speak, this seems like magic.

Thanks to the fact that LLMs can call tools, they can go to a particular website, fetch the data, clean it up, structure it properly, create absolutely gorgeous visualizations, and then explain the data.

There’s also the fact that these LLMs have absorbed a staggering amount of context from the internet, books, code, papers, documentation, and god knows what else. Which means they often have more context than I can possibly ever have about a particular topic. They can use that context to explain the data better than I could have done on my own.

And the result is all the visualizations you see on This Indian Life and Hope and Despair.

The other unique affordance they’ve offered me is that they’ve saved me the pain — the excruciating pain — of having to browse government websites, download CSVs, clean them up, and figure out what the hell is going on. You just point them in the right direction. In a lot of cases, you don’t even have to point them in the right direction. They’ll know which website to go to, where the data is, how to download it, how to clean it up, and how to make it usable.

I forget the exact number of articles, but combined together, it’s more than 60 articles across both websites. All of this was done in a span of a few days. It barely takes a few hours to create a really high-quality explainer after pulling a whole lot of data from different open databases like Our World in Data, the World Bank, WTO, government websites like MoSPI, and so on.

So this is a unique affordance.

Again, if your benchmark is not that of a normie, if you are a skilled data visualizer and a skilled programmer, this might seem banal. But to me, my benchmark is that I’m a normie, and this seems like magic. The affordance it offers is the ability to translate my vague desire to create data visualizations into an actual product, an actual site, so to speak.

And I’m having more fun than ever.

Going back to the point about fun, I recently had an opportunity to record a podcast with Karthik S, who is one of the best data visualizers I know. He’s an actual data analyst. He’s an actual quant. Unlike a lot of people who merely describe themselves with labels, he actually is all of those things, and has been for much of his life. He also writes No Enthuda, which you should absolutely subscribe to.

Even he was saying he’s having more fun. He was initially skeptical about LLMs, but now he’s having more fun than ever building a lot of things, both personally and professionally. And I kind of agree.

In a weird way, fun is a new affordance offered by large language models. As long as your interests have a software-shaped expression, which, again, just to not belabor the point, can leak into the outside world.

Because after I started publishing articles on This Indian Life and Hope and Despair, a lot of people started reaching out to me saying that these things were helping them understand a lot of things about India and the world.

The other thing I was thinking about, which was anyway my default prior, but which became clearer after speaking to Karthik, was this: the best way to get a sense of the affordances offered by large language models is to use the bloody damn thing.

Just get a subscription to Claude Code, Cursor, OpenAI Codex, or whatever else you like, and start using the damn thing. Put these tools to the test. Push them. Notice the patterns. Notice how they work. Notice the edge cases. Notice what they can do and what they can’t do. Speak to them. Start doing things.

You’ll be really, really amazed at what they can do for you, especially if you’re a normie like me.

And if you can’t afford the costly subscriptions to these tools, there are increasingly cheaper Chinese model like Kimi, Qwen and DeepSeek that are good enough for a lot of things. We live in a golden age where you can just bloody describe things and do things, and a lot of people are just sitting and discussing random things about, “Oh, LLMs are this, LLMs are that, they are just the median mediocrity of humans, they are just stochastic bullshit.”

Who gives a fuck?

They can do things for you. They can create unique tools for you. They can solve a lot of problems for you. They can help you have fun. They can help you build things.

Isn’t that enough?

Why do we have to get into pointless philosophical debates about the metaphysics of LLMs? Whether they are conscious, whether they are sentient — who gives a fuck? These are debates for scientists and philosophers. Let them have it.

Does it make any difference to you? When I say you, I mean normies like you and I. Absolutely not.

The irony in all of this debate about large language models is that people say LLMs are just vomiting the median mediocrity of humans, but a lot of people are worse than that. They don’t use LLMs as much as they should. They don’t put them to the test. Instead, what they do is listen to some loudmouth, mouth-breathing moron, and then pass off that loudmouth’s opinion about LLMs as their own.

In that sense, statistically speaking, they are below the median mediocrity of large language models. In that sense, large language models are as good as Einstein compared to the average mouth-breathing moron who just copies other people’s opinions about large language models and passes them off as his own.

Not only does he pass them off as his own, he passes out those opinions not really from the opening in one’s head — the opening whose purpose is to transmit speech — but from an opening in the hind part of the lower abdomen, which is not really meant for speech.

Look, there are real questions. There are real risks to using LLMs.

There is the question of cognitive atrophy if one over-relies on LLMs. What does it mean for one’s quality of thinking if one decides to outsource a whole lot of things to large language models? Where does this sort of cognitive debt, this cognitive cost, show up?

What does the rise of LLMs mean for society? There are other important questions too, sure.

But it’s not helpful to anyone to reason in absolutes without actually understanding these things, using these things, and then forming an opinion. That is a form not only of intellectual laziness or dishonesty, but a kind of epistemic stupidity that is not worth displaying publicly to the entire world. You might as well just tattoo the fact that you’re a moron on your forehead.

In a very loose sense, all technologies have upsides and downsides. And I don’t know, to my average normie brain, if these tensions have ever fully resolved themselves.

There’s a famous story about Ernest Rutherford and atomic energy. The precise details matter, so the line has to be used carefully. Rutherford dismissed the idea of practical atomic power as “moonshine” in 1933. Shortly after, Leo Szilard reportedly read about this and conceived the idea of a nuclear chain reaction. About a decade later, that line of thinking helped lead to the bomb.

So these debates are timeless. But for that matter, so is the stupidity about technology.

It’s our responsibility, to actually use these things and be okay with the tension of never having an easy opinion about all of these things.