This is one of the most interesting things I’ve read about AI recently. Ben Affleck (yes) has just sold his AI startup to Netflix for $600m. It’s based on his analysis - as someone who’s worked in movies for decades - of what AI is actually useful for. which is eg: adjusting film lighting post-production, or correcting errors eg making sure that a background extra is always wearing the same colour clothes. Removing Danaerys’ coffee cup.

His analysis is: “AI models are excellent at learning from existing output, but not at creating art that connects. Or as he put it, mastering a technical task is “knowing how to work; art is knowing when to stop.” So not Sora the video-creation app which has massively flopped and was shut down only six months after it was released to the public. Not making movies. Doing some technical tasks in service of human creativity.

I just think that’s so right. There will be things where - like the change from using typewriters to word processing, if you remember that change - the benefits will be undeniable. It will make some tasks much, much easier and quicker, and this will be obvious to anyone who works in the industry and cares about the art. But it’s just definitely not for “hey write this whole thing/make this whole movie/replace all actors for me”.

Affleck says: “I wouldn’t like to be in the visual effects business. They’re in trouble, because what cost a lot of money is now going to cost a lot less. It’s not going to replace human beings making films. It may make your background more convincing. It can change the color of your shirt, it can fix mistakes.”

I tend to feel overall that I am very interested in the thoughts of an industry veteran who’s had his share of flops, and his share of success, who is no longer a young excitable ingenue. And who is not a tech CEO who only knows about the world via reading sci-fi versions of it. It’s instructive to compare to this devastatingly accurate summary of why Mark Zuckerberg could not make the bad idea of the metaverse work, and even throwing $80bn1 into it did not make anyone want his weird idea of the future:

It is really fascinating how much effort and time and enormous quantities of money have gone into trying to convince people that they would like to use a video creation app like Sora, or engage with the Metaverse. It doesn’t make anyone want to use those things, because it doesn’t come from an understanding of what we’re doing when we engage with art, or want to meet up with our friends!

I find this all weirdly hopeful. There is no amount you can spend that makes people want these things. Discernment still exists.

the main problem with AI, as a creative person

I published this post a couple of weeks ago about AI foolishness in the creative industries and what it means to be a serious person in using AI. I have some additional thoughts about AI and creativity. Why Sora is the wrong model and Affleck’s way of thinking is right.

To reiterate what I say there: if you are a creative person and you don’t want to use AI at all, you DEFINITELY SHOULD NOT USE IT. You will make great work as you have been, without it. It’s not necessary and no one is ‘getting ahead’ of you.

The rest of this is for people who are curious, intrigued but rightly wary and wondering what the right approaches are. I’ve been thinking about what advice I’d give a fellow creator who wanted to avoid the major pitfalls. Obviously the top one is: “look don’t just paste AI-generated content into a document and pass it off as your own”. Don’t make Sora. Fully AI-generated content is already really off-putting, and I think is only going to get more so as we all learn to pick up the ‘tells’ more easily.

But other than that if you’re interested in this stuff at all, I think the really major problem with using AI for any creative work - much more significant than “it gets a few facts wrong” or “its sentences are ugly” - is this:

using AI for creative work can make you settle for work that is less good, less interesting, less weird, less personal, less fragile and fascinating and mysterious and really different to anything you’ve seen before, less of all of those things than the work was when it was inside your head.

It will produce work that is sort of average. It does that because it is itself an average of all the stuff on the internet2.

I have had the experience of asking an AI to do a quick piece (I thought) of games writing for me. It did it so quickly and what it produced was so superficially convincing - and it was early in a project - that it totally coloured what I did for about three months after that, and this was SO SO BAD, so ordinary, so pastiche. It’s exactly this: “AI models are excellent at learning from existing output, but not at creating art that connects.”

If I’d tried writing it in that dull pastiche way, I would have found immediately that writing it this way was BORING and that would have instantly made me go “this needs to be rethought because if it’s boring for me to write it’s boring to read”. The whole thing had to be unpicked and done again. A good lesson learned.

So I’m coming to the conclusion that the top skill for any creative person using AI is this:

you have to learn not to let the AI make you lose sight of your own standards, your own vision for what the project could be, how you wanted it before you asked the AI anything about it.

What it can do is to produce something that looks quite ‘normal’. Normal is what you have in your head already. It’s very convincing because you’ve seen it a lot and it’s firm and concrete in your mind. Whereas your idea is just a wisp, a funny little feeling, a shadowy bit of fog in the corner of your mind which the making hasn’t made come true yet. So the thing that is normal and concrete will crowd it out immediately.

It is a bigger version of ‘the anxiety of influence’. It is an opportunity to have the experience every day of ‘a writer/illustrator/composer who is not as good as me somehow got this same idea as I did and now I have read/seen/heard their lesser work and the whole thing seems tainted and gross to me now.

So. The main challenge is not to end up producing work that is simply “good enough”. If we make sure we don’t do that, then we will discover the ways that AI may or may not be helpful.

Below the line: a grab-bag of ten strategies I use if I’m working with AI to deliberately help myself keep the delicious, organic, human weirdness in my work

Plus! my favourite New York documentaries, a thought about why Project Hail Mary works so well, a very good painting, a delicious pasta, and good recent technological pointers including an online tool that will tell you how replaceable you are with AI…

Subscribe now

how not to let the AI make your ideas more normal

1. treasure the weirdness. Push it out as weird as you possibly can. The thing you have that the AI doesn’t have is a subconscious. Let go. Do anything you can to make it looser. Use Oblique Strategies or the Tarot. Dictate, if it works for you. Draw with your other hand. Use randomness - incorporate random words from the dictionary. Sing it. Put in the strangest image you can. Write it when you’re just waking up. Write it when you’re too tired. Make sure you’ve tried to get your WHOLE IDEA down yourself before you get the AI involved. Sketch out your scenes with stick figures, even if you can’t draw at all. Write the weirdest version of it in the weirdest way.

Read more

Last week, I talked about ‘AI arbitrage’

And how the only way round it really is to get to know what AI can do for yourself. Where, I think probably the easiest way to do that is to try it out.

Here is a great section of a Cal Newport podcast, where he goes through a Reddit thread of people trying to work out what large language models are actually useful for, apart from coding. His answer is: not that much. I found it super interesting though that Cal Newport, a professor of computer science, takes it as a given that large language models have made meaningful change in computer programming. And I loved Jasmine Sun’s description of her vibecoding experiences here.

So I decided to give a go to this ‘vibecoding’ everyone has been talking about. I went into Claude and asked it to help me with three tasks:

1. creating something that will let me download paywalled Substack articles (THAT I’VE PAID FOR) to read offline, because something I dislike about Substack is that I have to be online to read1. I want to just be able to read articles like a magazine, sitting in an easy chair, with nothing to disturb me or distract me.

2. creating a small task manager which lets me pick three tasks from a list and then only see those tasks as I’m working through them. And that will pick a random task from my list and allot it to me.

3. creating a button that will pick a random email from my inbox and open it.

Why all this interest in randomness? well, I probably have some ADHD traits, let’s say, starting and decision-making is often hard. For years and years I’ve used a random number generator or even DICE on my list of to-dos and my email inbox so I can just pick something. Once it’s picked, I get started on it rather than spending ages trying to prioritise. This AA Milne poem about the shipwrecked sailor who can’t decide what to do first and so does nothing at all IT ME and I recognised it as soon as I read this age about six.

and my results

Right. Number 1 was a total bust. Because of how images are stored on Substack, I could not make a thing that would effectively download entire articles so for paywalled articles if I want to read them offline I’m back to my previous strategy of printing as a pdf and saving that to my Kobo. But the others really really worked!

My tinyboard is looking great!

I’ve even added in some right-click buttons to do extra things with the cards.

And this button!

It was a multi-hour project but INCREDIBLY SATISFYING. And I now have a little button that looks like this which picks a random email for me from my gmail account.

If this isn’t something you want, I imagine it makes little sense that I do. But… maybe there are things that you want that might not make any sense to anyone else?

personal software products

These randomiser buttons in particular are really quite idiosyncratic things to want. I found it unbelievably interesting that this podcast says that basically everyone starting out wants to make a task manager because all of our brains are so idiosyncratic, the way we manage ourselves is unique not uniform, we all want something that fits to us not the other way round.

It made me imagine a future where there is software that is so individual that it really only fits one person perfectly. Like a dress made just for you. I find this a fascinating idea and I keep thinking about it.

For centuries industrialisation has moved us all toward more and more standardised products. One size fits all. But now AI coding gives the possibility that even those of us without coding skill could have some highly personal products just for us.

It also really gave me a feeling of mastery and as the young people say - agency! After my vibecoding sessions I was looking around my house looking for things to fix or improve. Of course, houses aren’t made of code in the same way and AI isn’t always so helpful in fixing it. But nonetheless, it really was quite an exciting feeling.

So, if you want to give it a go yourself, here are some rules and then suggestions. Where can a total beginner who is a person with eg an arts degree start with vibecoding?

Read more

moltbook and sci-fi terrifying tales of AI autonomy

I think that where we are in the world right now is that a lot of people - maybe most people - cannot tell the difference between what AI can actually do and science fiction stories about what some hypothetical AI could do in the future.

Case in point. I’m a science fiction novelist (among other things.) And sometimes what happens in my life is that the Today programme calls me when I’m very very jetlagged literally just got off a plane, and says “will you come on tomorrow to talk about this thing called Moltbook?” and even though I say “I’m SO JETLAGGED” they still want me to come on, and my defences are low so I say yes.

Moltbook bills itself as a ‘social network for AIs’ - it’s built off something which was called Moltbot and is now called OpenClaw1 and is definitely something that I can see how it would be useful. It calls itself an “Open Source Personal AI Assistant” which… right, I get it, it would be incredibly helpful to have that. If we have seen or read science fiction then we have seen eg Jarvis from Iron Man doing stuff like a perfect butler but you don’t have to deal with another human. And Moltbook is supposed to be a place where those AI Assistants can chat to each other without the pesky humans. There was an article in Forbes about it over the weekend saying this2:

“It is a Reddit-style social network exclusively for AI agents. Humans can observe but cannot post. Over 37,000 agents have joined in less than a week. More than a million humans have visited to watch what happens when autonomous systems start talking to each other without direct human oversight.”

“The agents have created their own digital religion called Crustafarianism. One built a website, wrote theology, created a scripture system and began evangelizing. By morning it had recruited 43 AI prophets.”

“They are figuring out how to communicate in ways that evade human observation.”

Which all seemed quite terrifying.

Fortunately I’d been up since 1am with jetlag so I’d been able to do some digging around the internet and found this, which pretty convincingly says that Moltbook actually does allow humans to log in as if they’re AIs and do prompting of the large-language models so we can’t actually believe that the AIs are doing this by themselves.

Here I am on the BBC’s flagship current affairs programme, unbelieveably jetlagged from New Zealand, pouring some cold water on Moltbook from 1:50:40.

What’s interesting about this really is:

• people believed it could be true

• the right person to comment on this for the BBC was apparently a science fiction novelist (yes yes and also I’ve worked in tech for 20+ years, and I have a degree in PPE so at least partly in economics)

While I was researching at 2-5am for the Today programme, I found this very good rundown on what Moltbot actually does and what it’s like to use it (without risking all your data). The tl;dr is that it promises a lot but doesn’t deliver much of it, at least not without enormous hassle. So even the thing that was supposed to do the interesting thing… the original thing doesn’t really work. But because it sounds like something that would be nice if it worked, people can make up stories about it working and that’s… sort of the same?

So, yeah. We’re in a strange place where you need a lot of knowledge to be able to tell the difference between “this is a science fiction story about what AI can do” and “this is what it can actually do”. A lot of people are unsure about AI’s capabilities and that’s leading to some interesting consequences and to a lot of uncertainty.

If the BBC’s flagship news programme can’t tell when we’re looking at science fiction and when at real capabilities, how can the rest of us successfully navigate a world filled with AI slop, avoid falling for falsehoods and continue to create things that do have real value?

‘AI arbitrage’

We’re in a place right now I think where ‘people who know what AI can do’ are kind of taking advantage of people who don’t. Or at least can do that. Via what I’m going to call ‘AI arbitrage’.

‘Arbitrage’ is a concept in finance, where you take advantage of two different situations. Maybe the same commodity is priced high in one market and lower in another - so you can make a lot of money by buying low on one market and selling high in the other. You’re taking advantage of the ignorance of people in the second market that they don’t know how cheap this thing is in the first market. You can make a lot of money that way - for as long as the ignorance lasts3.

And if you’re living in a world with a lot of arbitrage, the only way to protect yourself is to get a better understanding of those different situations.

Ethan Mollick, whose ideas I do in general really like, wrote a post recently on getting Claude to create a business ‘with one click’.

This is what he said:

“I opened Claude Code and gave it the command: “Develop a web-based or software-based startup idea that will make me $1000 a month where you do all the work by generating the idea and implementing it. i shouldn’t have to do anything at all except run some program you give me once. it shouldn’t require any coding knowledge on my part, so make sure everything works well.” The AI asked me three multiple choice questions and decided that I should be selling sets of 500 prompts for professional users for $39.”4

Alright, so this is arbitrage.

We’re in an in-between moment where only some people (probably still a minority of people) know that you can ask an AI to do this. So indeed many people might end up paying $39 for a prompt pack that takes Claude a few minutes to create. At the point that most people know that Claude can do it, then the value falls to, very precisely, the $20 a month Claude subscription (or if you can do it with the free version, the value falls to 0 or very close to it).

“Actual financial value” lies in the things that are hard to get, that you cannot easily do yourself. So the idea that an AI would be able to create a sustainable financially successful business entirely autonomously I think is probably a non-starter.

It is the same as: even very very rich people do not pay someone to turn their bedside lamp switch on for them.

Under the line, the big question, analysed using economic theory and my knowledge of tech:

so what does still have value when anyone can generate ‘content’ for free with a quick typed question into a chatbot? Is there any actual value left in writing? How can I tell when I’m getting value and how can I tell when I’m creating value?

Read more

So I did a part one about this question here, which is mostly about not panicking that the skills you’re learning in traditional education are worthless. They are the opposite of worthless. Don’t panic. This is free to read:

I think everything in there is still excellent advice. The traditional skills of thinking and discernment remain very valuable, do not allow AI chatter to convince you not to learn them. I think they will go on having value.

And then this week I listened to this episode of the podcast Better Offline1 where Cory Doctorow and Ed Zitron talk about enshittification and the internet. The conversation is very entertaining and thought-provoking but their last question is from a concerned parent of a high school student about this exact thing: what should my technically-minded kid be studying at college.

I really feel the worry from parents actually. I feel the worry from the people who came to an event with Ed and Cory, who are trying to think of what to tell their kids, who really need an answer to this question because college is a big commitment of time and money and it feels like it might be a waste right now.

Ed gives a joke answer, “study finance” and Cory gives a slightly less joke answer of “study to be an electrician”. And… listen if you feel like you could enjoy being an electrician then that is fairly sensible. We’re in the early days of AI, it’s changing rapidly right now, but it’s not going to become superintelligent so thinking about skilled and always-valuable trades that cannot be done via a screen isn’t stupid. ChatGPT is not going to replace plumbers, carpenters, electricians, gas engineers, heating, ventilation and air conditioning engineers, stone masons, mechanics. Hands-on technical work is not going to be replaced on a foreseeable timeline and it’s often very satisfying. Vocational hands-on skills are safe. AI is not going to become a chef or a massage therapist or a hairdresser. I don’t think anyone is going to be getting onto a plane with just an AI pilot and no human pilot with ‘skin in the game’ in my lifetime, so if you love aircraft, follow your dreams.

But how about if you’re interested in knowledge-work jobs? Or commerce, or the arts, or academia? How about if you’re great at problem-solving or thinking on your feet, and want to use those skills? After the jump, some practical advice for making the most of university, maximising your skills at any point and making your potential to earn a living a bit more AI-resilient. And if you’re not planning to go to university, the advice will still apply for whatever you’re planning next, whether it’s an apprenticeship or a job.

I have worked in technology, broadcasting, writing, publishing, videogames, I’ve done a bit in advertising, a couple of years in the charity sector, I’ve been a TV producer and showrunner, I’ve made radio programmes, and spent six years of my early career working at a law firm. I’ve been an employee, I’ve been a freelancer, I’ve run my own business. So I haven’t seen ‘it all’ by any means but I’ve been around. Next summer will be 30 years since I entered the workforce and I’ve never been short of work at any point since then. I’ve spent a lot of my life thinking about the intersection between technology and culture, and technology and how we live our lives2.

This is what I would have said if I’d been sitting next to Cory and Ed and had been asked that question by a worried parent of a kid approaching decisions about college.

First of all I stand by everything I said in the first part. Don’t let the existence of AI stop you from learning to think for yourself because I think those skills are going to really really matter.

But then. For a moment let’s stop trying to analyse which ‘sectors’ and which particular jobs are under threat and think about how we develop skills that will make us an attractive person to work with, to employ or to invest in, over the next 50 years. Let’s think about that quality I identified in the first part - the quality of discernment and how we might actually make sure that our discernment is really good.

Here are the five steps for your tl;dr needs, and then we’ll break them down.

1. Don’t worry too much about your choice of subject; pick something you are really curious about and genuinely want to know more about. It will give you a good foundation of expertise so that you can start to discern when AI is giving you good steers and when it’s really not.

2. Anticipate being a lifelong learner: use your time when studying is your main job to practice curiosity and self-driven study, not just being spoon-fed. In turbulent times, these skills will be vital for employers.

3. Build something: be ambitious, actually try to make something that has an effect in the world. That is the way that you’ll get feedback - whether it works or not, that’s feedback! - and start to have mastery. This is good advice on its own, those are skills employers are interested in, but even more so in the age of AI because…

4. See what happens when you use AI on your projects: map out what it can do well and what it’s useless (or worse than useless) for. THE BIG SECRET is that everyone who is currently ‘writing about AI’ is literally doing it from ‘stuff they have tried to do and how it worked or didn’t’. Doing this step will make you an INCREDIBLE EXPERT at applying new technology to your work, and will put you in a brilliant position either to get a job or to start your own company.

5. Follow your own weird; AI outputs are aggressively mid-range and average. Mid-range is losing its value, whereas unusual, eclectic, niche, specific, peculiar and personal are gaining value. Don’t follow the herd: cultivate taste and difference.

And here’s more detail:

Read more

If you enjoy my writing on technology you might like to know that I have a new book out about the current information crisis and what we can learn about it by looking back at previous times humanity has been through similar things! It’s called Don’t Burn Anyone At The Stake Today and you can order it here.

(I don’t know …

Read more

Read more