I’d like to introduce you to TimBot. We’ve been working on him for a while now.
TimBot is our AI-driven media expert. His answers are based on the contents of my book Media Unmade, along with the last two years of Unmade’s archive.
Now, we’re ready for you to take a look. You’ll find TimBot at timbot.ai
Before you try TimBot, let me share some BIG caveats.
First: TimBot is a prototype. Sometimes he breaks. If he does so in an interesting way, please do send us a screenshot to Cat McGinn.
Second: Because TimBot’s computing power is driven by elements of both ChatGPT-3.5 and the recently updated ChatGPT-4, he can be slow, particularly when America is awake. If things are running really slowly, you’ll simply see an error message.
If you’re interested in trying him out, I suggest doing so when this email arrives, while America is still asleep.
This is a collaboration which has been led by the brilliant Nic Hodges at generative AI consultancy Fear of Missing Art, and inhouse by Unmade’s Cat McGinn.
Cat is curating next month’s humAIn - human creativity X AI conference, and will talk more about TimBot there. It’s not too late to buy a ticket to humAIn, by the way. The key principle for curating humAIn was that we wanted everyone on stage to be already grappling with AI, not just talking about it. It seemed important that we got our hands dirty too.
Incidentally, based on our fantastic recent experiences on this project, I’d recommend you talk to FOMA about your own AI needs. We’ve also been working with Abe’s Audio on this project. We’re hoping that by the time humAIn arrives, TimBot will be ready to speak to you too.
If you’re curious how TimBot works, we’ve tried to lean on OpenAI’s ChatGPT for what it does well, and avoid what it does badly. ChatGPT offers amazing computing power, but as you have probably heard, it can be prone to fantasise when it doesn’t know an answer.
Nic has designed TimBot in such a way that it surfaces what I’ve previously written as the basis of its information, rather than what’s on the wider web. That means the information he’s aware of is more narrow, but his answers should be more accurate.
In the main, this is a proof of concept with applications and implications for both book and journalism publishers. Can you take the knowledge and writing style of an individual writer and turn it into a useful chatbot?
The chart below shows what’s going on under the hood.
We use ChatGPT-3.5 to reformulate the user’s question, including pulling out relevant content from the book and newsletter. ChatGPT-3.5 then uses this to synthesise the best answer, which it finally passes along to the more powerful, but slower ChatGPT-4.
ChatGPT-4 then tidies up this answer and, based on my previous posts, attempts to recreate my own writing voice. As you’ll see, he’s chatty, but not entirely me, just yet. In time, I hope he becomes 20% less flippant and 20% more detail-oriented.
The process happens in a few seconds (and sometimes stalls). It's certainly not as fast as typing a query into Google but it still feels kind of miraculous.
I’m not the only person to have thought of using AI to create a distinct author voice. The idea that AI could be used to train a chatbot on a particular writer’s style has been occurring to several people almost simultaneously.
Marketing writer Seth Godin launched a similar offering to TimBot this week.
Godin’s bot moves at a similar speed and seems to go through roughly the same thought process as TimBot.
And listeners to Scott Galloway’s various podcasts will have heard him talk about how he is working on a similar reproduction of his writing voice.
Intriguingly, I spoke to a marketer a few days ago, who has taken the writing of Byron Sharp, Mark Ritson and Scott Galloway to create for himself a private virtual marketing adviser. Talk about a dream team.
For authors and publishers, that becomes an issue and challenge of who owns that content. That marketer I mentioned didn’t ask permission from any of the individuals involved. There’s a moral question there that copyright law may not properly cover.
We considered also training TimBot with my dozen years of writing for Mumbrella, but in the end concluded that although it would probably have been legal, it was an ethical grey area, as I don’t own that content, even if I’m the author of it.
The negotiations going on between OpenAI, Google and the major news publishers will hopefully create some sort of licensing framework that could apply to the training of specialised AI chatbots, just as the likes of the Copyright Agency has for content in Australia.
I could also see a time when any new non-fiction book gets an accompanying chatbot.
Another fascinating side issue is the implications of this sort of product for the development of quantum computing.
A major barrier to the utilisation of ChatGPT is speed. As the Large Language Models of ChatGPT-5 and upwards no doubt becoming increasingly powerful, the question is whether they can keep up with demand. Yet they’ll likely be too big to be stored locally on conventional computers.
That’s where quantum computing comes in (declaration of interest: I’m an investor in an Australian quantum computing startup). If that technology becomes ubiquitous then we’ll see the likes of TimBot responding in fractions of a second, not half a minute.
Even with the limitations of speed, some of TimBot’s output has been uncanny. On occasion he has made connections between items I’ve written that I had not spotted myself.
And at the very least, he’s become a useful tool that saves me turning to the index of Media Unmade when I’ve been checking a reference or date.
Please do give TimBot a try. And remember, if you don’t feature in the book, or haven’t been written about on Unmade, TimBot is unlikely to be able to tell you much about yourself. Please do tell us what you find.