The robots are coming for us!

I’ve always been fascinated by the overlap of technology and creativity; whether that be an interest in 3D photography, the creation of Lego stop-motion animation videos, or even my dissertation (which had the grand title of “Algorithmic harmonisation of a melody into a four part barbershop arrangement). But here we stand in March 2023 and it feels like the world is at a pivot point. Why? AI.

Before I start to unpick this, just a note – it’s called AI (Artificial Intelligence), but I really don’t like that phrase, because of the science fiction connotations. In my dissertation nearly 20 years ago, I instead talked about neural networks and genetic algorithms. Much of the development since then in the “AI” space has been using the language of machine learning – using training sets of data to build probabilistic models (i.e. billions of paths through a network of possibilities).
But let’s not get too far into the tech – instead I’ll focus on why I’ve said we’re at a pivot point.

Back when I was visiting universities about to study computer science, AI was a buzz word (or should that be buzz acronym?) and there’s been slow and steady progress over the last 20 years, but the cost for individual use has been incredibly high. Huge costs of training these models and not quite so large costs of asking questions of them, for an output that was… OK. What’s changed in March 2023 is that suddenly the output is more than just OK, and the cost of querying them is low, and the compression on these LLMs (Large Language Models) is such that some can be run on a personal device.

Some of the prompts I can feed to these LLMs are:

• Re-write the first verse of “S Club Party” in the style of a Beatles song
• Summarise the Gettysburg Address as a limerick
• Tell me how the Spanish Inquisition is relevant for a delivery driver today
• Write some php code that will grab the top headline from the BBC website once an hour

And you know what… it’s pretty good. There’s no “intelligence” as such, but if you give one of these LLMs a prompt (just a sentence in plain English as above), it will word by word (well, technically it works on tokens… one token is generally around three quarters of a word) try to work out what the most likely next word should be. They have been trained on vast swathes of data to try to understand and model the patterns and relationships between words, that are then transformed into probabilities. Basically, predictive text on steroids.

When I was originally going to write this blog post, I would have said “and these LLMs aren’t attached to the internet, so don’t know real live facts”… but just last week OpenAI announced the concept of plugins, including one that can load stuff from the internet – this stuff is progressing very quickly.

So… how do you get access and ask these questions of an LLM? The simplest way is with OpenAI (other LLMs are available!) – visit chat.openai.com, create a free account, then you can start asking questions.

But I wasn’t content with just asking questions – could I build something with this? Oooh, they have an API. Now, given that all of this is based on training data, what might feature heavily in there, that could also potentially provide something interesting, thought-provoking and possibly helpful – how about something to do with the bible? So, I created an API key and, with my free $5 credit got to work.

My initial plan was to build a website where a visitor could enter a bible passage and their occupation, then the site would provide a summary of the passage relevant to their occupation. But, maybe that’s not a good idea giving free text fields that might be open to prompt injection (i.e. malicious prompts) and with a cost per question, it could get expensive for me really quickly. I then moved to the idea of possibly loading up a bible passage of the day… but then that could also get expensive if it was regenerating every time.

OK, how about setting it so that when the first person visits the site in a day, the code makes the call out to OpenAI, then saves the result into a WordPress blog post… which others can see. Remember that last question from above where you can ask the LLM to write code for you? Well, I had a morning free to write all of this… so we got to work… with a lot of the code written for me. In the end I stuck with 4 bible passages a week taken from the Church of England lectionary, as I couldn’t find a suitable site that lists a bible passage of the day.

And here is the result: onebread.co.uk. The initial prompts ask it to describe a visualisation of the passage, summarise it as a limerick, provide some basic action points, and related passages… but ultimately there are many things that we could do with it:

• Generate an image for each passage using Dall.e (I intend to add this to the site over Easter)
• Ask it to identify a craft activity associated with the passage
• Ask it to explain how these passages are related to the top news headline drawn from another website
• Ask it to re-write the passage as a parable / with a particular metaphor

The only limit of this is the creativity of the prompt writer… and it’s not just LLMs with text-based interaction, there’s also generative “AI” that will create images (Dall.e referenced above is just one of them), there are models that will generate videos, models that will generate audio, or translate audio to text. And all of them are now at the stage where actually, they’re really good.

I honestly believe that we have pivoted from the Information Age into the Intelligence Age. It is a tool to be harnessed and will likely change the way that we use and view computers. It will also likely change a lot of jobs, challenge assumptions about creativity and change the way we look at things – how can we trust that something is real, rather than created by a computer?

I was going to end this blog post with “What do you think?”… but actually that question could just be fed to an LLM to answer… so instead I’ll change the question to “How do you feel about this?”.