Categories
Uncategorized

AI & Books: notes from a moral panic

On July 21st, 404 media published AI Companies Are Buying Tons of Old Books Because They’re Free of AI Slop. It describes the practice where, to train their models with ‘premium’ human-written data, AI companies are using brokers to buy copies of books they don’t already have in their datasets, which they scan using a destructive method (they’re not allowed to keep a physical copy after scanning due to copyright law).

On July 25th, Futurism repackaged this article as AI Companies Are Buying Antique Books, Ingesting Their Contents to Train Models, and Then Destroying Them at Incredible Scale. The books have now become “antique”; being generous, I could put this down to a misunderstanding because some of the books come through antiquarian booksellers, rather than being antiques themselves.

One book retailer involved in these sales, Half Moon Books, actually describes the titles that were ordered as:

Not what I would consider rare or antiquarian. They were more obscure academic commentaries on philosophy, politics and science. Nothing signed, no first editions and nothing that hadn’t been published in the 20th century or later.

Being less generous, I note Futurism’s use of scare language about the “incredible scale” of the destruction.

If we’re talking incredible scale, the book industry pulps hundreds of millions of books per year (we don’t know exactly how many because they never release that information). In a 2025 article, Pulping Fiction: Where Books go to Die, Alexiane François writes:

Every year, millions of unsold books around the world are destroyed. It is hard to know how many books exactly, as there is a huge effort from the publishing industry to hide the scale of the waste. But in the U.S. alone, estimates hold that between 65 to 95 percent of returned books are destroyed. This does not happen only in the US. In France, the volume of pulped books represents around 140 million copies every year. 

In the most recent available figures I could find, the UK pulped 77 million books in one year. But that year was 2009, and nobody seems to have talked about it since then.

Anyway, four days after Futurism’s piece, the story starts making the rounds of social media, with all the subtlety and consideration for the truth that you’d expect. This is probably the most egregious example:

AI companies are buying up the world’s rarest books, scanning them, and then destroying the books so there are no copies. This is one of the biggest alterations of history we have yet to see.

This is a HUGE exaggeration of the original reporting. Every one of the claims I highlighted above are false. “Old” books have become “the world’s rarest” books. “Destroying the physical copy during scanning” has become “destroying the books so there are no copies”. A report of a practice that some people find morally uncomfortable has become a nefarious culture-spanning conspiracy theory.

I’d dismiss it, but it’s been liked over 56,000 times already. Some people like to claim that AI is unreliable because it can hallucinate but honestly, people are capable of far worse.

It’s noted in the initial article that, on the website of one of the brokers, it says “The optics problem is real. ‘AI company destroys two million books’ is not a headline that generates sympathy”. It may not generate sympathy, but it certainly generates bullshit.


Aside: when I posted about the scale of book destruction that already goes on, a typical reaction was “yeah, but we’re mad because this destruction is to train AI” (I had an excellent example of this, but they deleted it). This is a moral judgement, not a rational reaction.

Andy Masley wrote an excellent article, Toward Liberal Environmentalism, which goes into much more detail about this type of criticism. I urge you to read it in full, but this is a very salient part:

We the good people broadly know what is good and what is bad, how scarce resources should be spent to maximize people’s wellbeing, and which contentious ethical values are right and wrong. Because we’re in an environmental crisis, our main goal should be to make sure no resources are spent on the bad things, since any environmental cost paid for something bad is by definition pure waste, which we just can’t afford right now. We have no responsibility to mediate the complex, contradictory desires of the billions of people alive right now, or to think about how to trade off their desires for radically different goods and ways of living. Instead, the goal is to attack the bad guys, the people spending scarce resources on things we already consider bad for other reasons. All good people know what’s bad, or will once they’re awakened from the lies of the bad people.

“I don’t like AI so this book destruction is bad book destruction.”

Categories
Uncategorized

How I use AI in creating my newsletter

For a few years now I’ve written a weekly newsletter, Tech Landscape, that rounds up new feature and product announcements in digital technology (90% AI these days), social media, and online culture, with some comment and insight where appropriate. It’s broadly aimed at people who work in or around digital marketing, or just want to stay up to date with what’s going on, and is pretty useful, I think.

But the increased pace of change in the last few years means I have a lot more to cover than I did when I started, and it takes up more of my time. So recently I’ve been looking at ways to use automation and AI to help me with some of that burden.

The Words

Producing the written content is a balancing act; I want to automate to save me time, but I don’t want to automate too much because then it loses any value. Here’s my current workflow:

  1. Throughout the week, I save all the links in Airtable
  2. At the end of the week I paste all of the links into a custom app I built, which uses Gemini to crawl and summarise them in an approximation of my voice.
  3. Some links (especially social links) are blocked to crawlers, so I write those myself
  4. When all the summaries are generated or written, I paste them back into Airtable
  5. I export from Airtable (as a CSV file) and import it to another custom app I wrote which formats it for the newsletter
  6. I post that into Substack, then do a few rounds of editing, writing, and rewriting

Those final rounds are critical; that’s where my voice shows up. I write all of the comments and asides myself, but also the automated summaries are rarely perfect and I almost always have to make tweaks or changes. That could be to make a sentence more human, to change American spellings, or to focus more on the things that I think are important.

For example, this is how Gemini summarised a recent story:

Adobe announced a definitive agreement to acquire Topaz Labs, an AI company specializing in advanced video and image enhancement models.

And this is what I edited it to:

Adobe is buying Topaz Labs, the company that specialises in video and image enhancement and upscaling.

Whereas this one I had to rewrite entirely; not because it’s wrong, but because I think it focuses on the wrong features:

Figma summarized new product updates from Config 2026, an update that introduces a built-in timeline animation feature, AI-generated shaders, and prompt-driven custom plugins.

Here’s my final, more detailed version with a more appropriate emphasis:

Figma announced several product updates at its Config 2026 event. The headline was code layers, which makes every layer on the canvas editable with code, unifying design and build in a single interface; it’s coming soon in closed beta. Figma Motion is a timeline editor for the canvas, enabling motion design of every element. Figma and Weave are becoming more tightly integrated. And the design agent is coming to all users, and now lets them code their own shaders and plugins.

I reckon my current semi-automated process saves me an hour or more each week, but maybe 75% is still hand-written. My next project is working out ways to make the process much simpler, while bringing that writing ratio down to about 50%, but still preserving my voice.

The Cover Image

I’m also trying to get a better grip of agentic workflows and MCP, so I paid for a Manus subscription and set myself a starter task: to create a cover illustration. These were the steps I settled on:

  1. Check the Airbase table [via MCP] where I store each week’s stories, and get all of the written summaries for each story.
  2. From those, synthesise an over-all cover concept, research [via the Web] current design and illustration trends, and turn that into an image prompt.
  3. Send the prompt to Magnific [via MCP], choosing the image model best suited to the task

After testing and revising it a few times, I turned it into a repeatable Skill. And here’s the result, generated with Reve 4.1:

A pair of human hands and a subtly robotic hand collaboratively weaving a luminous, intricate network of glowing digital threads and physical yarns on a dark background. The threads form a complex, dreamlike tapestry that blends smooth digital light filaments with tactile, embroidered textures, patchwork fabric swatches, and hand-stitched details. At the center of the weave, a radiant node pulses with soft electric light, symbolising an AI agent. The composition uses twisted geometry and gentle optical illusions, with the threads forming recognisable shapes: an XR headset, a film frame, a social graph. Color palette: muted psychedelic tones — dusty rose, sage green, warm amber, and electric indigo — against a deep charcoal ground. Mood: contemplative, warm, quietly surreal.

Clearly it isn’t great; it’s a bit overly-literal in some places, overly-metaphorical in others, needs more of a high concept. But I’m learning a lot about how to prompt multi-step workflows, which is useful, so I’ll revisit this in the future.

Categories
Micro Blog

Taste is the bar of quality in the absence of correctness

In math, you know 2+2 is 4. But how do you judge a great article? A great pitch deck? A great painting? A great design? Much harder. So much harder that people have boiled it down, for lack of a better term, to "taste." But taste is nothing more than a rule of thumb: compressed judgment learned from pattern recognition, from exposure, from being opinionated and having the courage to say yes to certain things and no to others.

— Thais Castello Branco, Stop Debating Taste

Categories
Uncategorized

The Future of AI Images

How a model of the world can make a more believable breaking wine glass

Last year I thought that AI images had more or less hit a plateau in terms of what they could offer. Like the quality was, in the best of them, as photo-realistic (or paint-realistic, illustration-realistic, etc) as they could be. There were some problems still to be solved, such as control over composition and pose, but I didn’t think there were more breakthroughs to be had. I thought we were basically done.

I was proved wrong by Google’s Nano Banana Pro. Its breakthrough was to integrate image generation with the reasoning of Gemini 3, so you can create much more exactly what you want. You can edit with your words, it preserves image states much better, you can use Gemini to ask it to “fill in the blanks” of your prompts… basically it reads your intent and ‘understands’ you better. It creates much more of a close relationship between you and the model; you don’t have to know what to ask for in its language, you can use your language and it gets it really well and is able to do very complex outputs including text and infographics.

That was a breakthrough I hadn’t foreseen, and it made me wonder what else I was missing. What’s next for image models?

The Falling Wine Glass Problem

I think it comes down to this: if you ask an image model, no matter how advanced, to generate a picture of a wine glass full of red wine hitting the floor and spilling wine it will give you all of the ingredients visually, but assembled incorrectly. The physics will be wrong; the wine will already be on the carpet or will be spilling out of the glass in an unrealistic way.

Six images of a glass of red wine falling on a white rug; all of them are wrong in some way
Prompt: A glass of red wine falling onto a white shag rug, captured the millisecond before impact.

Image models create visual mimicry; they know what elements should be in the scene, but none of the actions happen in the correct sequence (except perhaps by luck). And so I think that’s the next step: causal understanding. Cause and effect, action and reaction.

Changes Over Time

Maybe image models need to ‘imagine’ before they generate. NVIDIA recently released an experimental image model called ChronoEdit which uses temporal reasoning; when given an image and prompted for a change it will (kind of) imagine running the changes forward through time, like a video, to reach the end state and return the final frame as the image result. For example, given an image of an empty room and asked to “insert a metal object ball with smooth curved surfaces into the scene” it imagines the ball rolling into place and then shows you its final position.

This is a quite slow and intensive approach, but its advantage is a result which should be much more consistent with physics as we know and observe it.

But this ‘imagined video’ approach isn’t a perfect solution because video models themselves struggle with causality. As with images they’ll get some things right to a certain degree but they’ll also get things very wrong — like the falling glass.

Although advanced models such as Veo 3 have already exhibited some learning of physics — for example, the difference in reaction when dropping something light or something heavy in water — they often get confused in complex scenes where there’s more than one thing happening. They might ‘know’ the gravity of a glass falling, and a bit about how fluid should move, but they still struggle with the sequence: the glass falls, it hits the ground, the liquid explodes out. They don’t understand that one thing causes the next.

Enter World Models

You may have noticed that recently I’ve been posting more about world models — that is, AI models that are trained (largely on video) to the extent that they begin to ‘understand’ the physical laws of an environment, such as a video game or — ideally — our reality.

Some world models only generate a place, a persistent 3D space that can be moved around in, but not interacted with. But with highly complex and powerful training, they can generate an experience, a persistent space that can also be interacted with; they can learn not just what motion looks like, but how it works. Rather than 2D mimicking 3D, they create a latent space in 4D, with some understanding of motion over time. Probably the most advanced in that area at the moment is Google’s Genie 3:

The experiences it generates look quite simple — a character moving around in a world space — but have some understanding of physics; for example, if you make a character walk towards a wall it will stop and be unable to pass through it because that breaches the laws of physics. As the technology progresses, future versions will hopefully learn more physical laws, and cause and effect too.

World models have lots of potential uses, from prototyping video game levels to training cars and robots. And they’ll clearly be useful for generating videos, but will also be very good for generating images.

From Visual Mimicry to Causal Understanding

Returning to my earlier example, if I prompt for a picture of a glass of red wine falling on the carpet and spilling then a world model should be able to “imagine” it using its understanding of causality and then generate an image that’s essentially a frozen moment in time, depicted in a way that makes sense to us. That might be slow and intensive, but these things tend to become faster and more efficient over time.

So that, I think, is the next potential breakthrough for image models: a baked-in understanding of how the world works, its physics and causality. Rather than being just a visual representation of a scene with all the ingredients but in the wrong place in the wrong order, everything will just look right to us.

I don’t know when this is going to happen, but I imagine we’ll start to see it more this year.


Post-publishing updates

PhysicEdit is a new experimental approach to physics in image models, using a version of QwenImage Edit fine-tuned with 38,000 videos of physical state transitions.

Nodes Aren’t the Future of AI Creation. Here’s What Is. Bilawal Sidhu on the shortcomings of current creative tools, and the approaches that we’ll likely adopt in the future.