We all know Henry Ford’s apocryphal saying “If I had asked people what they wanted, they’d have said a faster horse”. I suspect there are a lot of ‘faster horses’ being built with AI. No shame in that, we are all constrained by the limits of our training set.
Erik Brynjolfsson of Stanford University asks us to imagine what glories the ancient Greeks would have requested if granted access to modern engineers? Automated goat herding? Clay pot factory lines? Both labour saving, but they’d fail to imagine the staggering wealth of the modern world.
Even the greatest thinkers would first automate existing tasks rather than create truly new wealth. Image by DALL.E3
Brynjolfsson goes on to describe the ‘Turing Trap’, the temptation to build machines so effective at emulating us that they simply supplant us. If such a process were rapid then it would invite dystopian wealth transfer not utopian wealth creation. We’ve all seen that movie.
Industrial Inspiration
There’s another analogy which points to a way out of this trap. Near my home is the world’s first iron bridge, dated 1779. It is designed and constructed as if made of wood, because no one knew how to design an iron bridge. As such, it spans a distance equally achievable with earlier materials. Years later, Thomas Telford understood that iron’s super power was strength in tension, he built the world’s first substantial suspension bridge at the Menai Straits, nearly six times longer than the iron bridge. The revolution had finally arrived.
The iron bridge in Shropshire, designed in the style of a wooden bridge, 1779. Photo Graham Hobster
Think of Super Powers
So, to stimulate our imaginations, what superpowers can AI augment us with? I’d like to start the bidding with the following six. Get involved! What superpowers am I missing?
1. Converse at Scale
Until recently all conversations were 1-to-1. But AI can interview and interact with thousands simultaneously, revising understanding iteratively. In the 1980’s newspapers produced one product for millions of consumers, now we tailor the media product to each consumer. Questionnaires may go the same way (Trove).
The first corollary of being a super power is that conversing at scale is un-human. The film ‘Her’ (2013, Phoenix+Johansson) movingly dramatises the distrust which could arise.
2. Emotional Intelligence
Google’s AMIE has already been demonstrated to have a bedside manner as good as a doctor’s, but of course AI has the advantage of time to give. Meanwhile, Character AI has millions of teens pitting their conversations skills against bullies, manipulators, lovers and leavers. Is this ok? Again, this was anticipated by the movie Her, the scene with Amy.
3. Expertise at Scale
GPT4 is not cleverer than people in any single specialism, but it is the best read generalist around. Recently, I built AI teams to write software using Microsoft Autogen, at first I built the teams like people; a tester, a developer, a scrum leader. This was ok, but expensive; lots of chat. Of course, I had neglected that the AI has all those skills in one mind. It was more effective to create a team of two; a coder with a critic for self reflection. That worked elegantly.
4. Loop of Creativity & Process
For decades computers have been mischievous genies, precision was required or we’d get an error. Now, we can explore problems with creative and ambiguous language, ask the AI to interpret us and map that intent to rules, policies or code. Then iterate around the loop to reach a robust result. Or we can simply ask the AI to cycle around this loop with itself (see coding with Autogen).
5. Synesthesia
Synesthesia is where people experience a sound as a colour, or another blend of the senses. Multimodal AI systems like Gemini 1.5 Pro project their training data to a single space, blending images and text, achieving a skill most humans cannot. It can read text (e.g. code) for the images which that code creates, it processes the code as if it were the image. This superpower is closely linked to the next.
6. Comprehension at Scale
As Instagram shortens our attention span, AI increases its own. Gemini 1.5 Pro has a context window of 1 million tokens, approx 7x the entire series of Harry Potter books. My personal ‘context window’ is likely no more than a chapter before I start losing detail, certainly no more than a book. What perspective can AI deliver for us when reasoning over such a wide expanse of information?
Well, we can load a small company’s entire code base and ask the AI to code with it, or propose new functionality. No need to fine tune or retrain the model. But, I feel this is scratching the surface of the possibilities, an imagination failure on my part.
Conclusion
These are just my suggestions, an attempt to escape the confines of my conditioning in a world before AI. That past world constrains you too. If you have suggestions on other superpowers, or mechanisms to escape these self-imposed constraints of yesteryear, then do comment below!