That’s it, folks, AI version 1’s a wrap
Well, we’ve done it. Pop the corks, and let’s enjoy a socially distanced Zoom party to celebrate.
The current generation of AI, and in particular machine learning, has ended its beta phase and is now officially done. We can look forward to incremental update releases over the coming decade, but it’s packaged, available, everywhere, reliable, and powerful. It isn’t intelligence, of course, it’s mimicry and extrapolation, as well as advanced analysis of complex data sets. So more statistics than smart, but nonetheless, it’s pretty amazing. There’s a Simpsons episode where Bart, I think, asks “zoom and enhance” and Lisa replies “Bart, you can’t do that, if the data isn’t there, you can’t create it” and promptly moves his head closer to the screen. But now you can. From a rubbish pixellated image of a face, it’s now possible to magic up a likely detailed high-resolution face using machine learning. And it’s impressive. For law enforcement, this is going to be a powerful tool in their box (and I’m sure there are some less attractive uses, too, but c’est la vie when it comes to technology).
In most of its chosen application areas, so long as it is managed well, it delivers well. In others, it’s surprising how quickly it breaks down. Take this little dictation on an iPhone recently. I’d like to believe that I speak relatively clearly, so this is particularly amusing. Top is what I meant to say, which is an early draft paragraph in a blog post I am writing for later this month, bottom is what Siri heard. If you can link the two together, you’re better than I am.
What I meant to say:
“These are blitzscaled businesses that have operated at a loss in new territories in order to get a foothold, driving the small local players into submission or out of business. Where that does not work they have aggressively acquired the competition. This comes with a cost, and if you do not rapidly achieve a monopolistic utility position it cannot be sustained.”
What Siri heard:
“The scale businesses that have operated in order to get a foothold in homeThe snail businesses that have operated lost in order to get the bus home and where that doesn’t work they have aggressively a client of this message it’s going to come to that because if you do not achieve I’m not ballistic utility position and it cannot be sustained long time whose bed looks engine the pig is well.”
I particularly like “in order to get the bus home”, and that thing about engine-the-pig. I’m really not sure where either came from, it was 6:30 AM on a Saturday morning, on a brisk walk. Absolutely silent, so it wasn’t a challenging environment for speech recognition. Modern AI does not understand, it does not comprehend. So it can’t correct for context: and this is why autocorrect on Apple devices doesn’t correct “fir” to “for” effectively, but it will present a cute little fir-tree emoji as a suggested word. Some are better than others, with massive databases used to link words together according to proximity frequency, and those would see that “It’s” and “you” are rarely connected to “fir” but more connected to “for”, given the edit distance and sentence structure.
So much more to come
Whilst we can look forward to constant updates to this generation of AI, and the occasional new feature and application to surprise us, pretty much nobody believes that it will lead to artificial general intelligence: thinking machines, that can out-compete humans on things like getting dressed, making dinner, assembling Ikea furniture, understanding irony, sarcasm and humour or passing my self-driving car test, The Courchevel Challenge. Most experts in the field have long since accepted the limitations of the current approaches on this journey and are now even daring to talk about it in public, on the basis that admitting it no longer scares away the money, as version 1 is—and will continue—to be a great value generator and one that is usefully improving our day-to-day lives.
As collective learning (such as the work we’re doing at Fetch.ai), scale-through-decentralisation and the increased prompt availability of vast data-sets feed further in, we’ll see significant improvements in what all this stuff can do. Dimensional reduction, for example, using recurrent neural networks, can provide a semantic proximity that greatly aids fuzzy search-and-discovery in the autonomous agent space. Think of it as a sort of encouraged serendipity for the digital world. One way or the other, few aspects of our lives will be untouched by what this version leads to. From healthcare, through mobility to optimising the utility in what we have, systems that learn, both in real-time and are pre-trained, are the ubiquity of the next decade.
It’s a wrap. What’s next?
And so there we are. It’s a thing. It’s there. We’re done, and we’re moving on. Minor versions, major versions, it’s all coming but it’s based on the same root. Let’s face it, Microsoft Word is Microsoft Word, even though it is an entirely different experience than it was when it first launched in 1983. It’s still a word processor, and whilst it won’t write your work for you, it’ll—mostly—point out your awful grammar and terrifying spelling. And, of course, the scale and beauty of what you can construct is light-years beyond what we dreamed of when we first fired it up to write a letter all those years ago.
So what’s next? Where and when will we see the grand leap, from the path that we’re on to an entirely new approach to machine intelligence? There are plenty of irons in the fire. Those that are revisiting biology as a source of inspiration, particularly interesting given it’s the only working example of true intelligence that we have. Then there’s the acceptance of just how important imagination is: being able to see possible futures, and the reward or punishment of them unfolding, without actually doing it. Our ability as a “what-if” machine is, it is fair to say, a vital foundation piece when it comes to being really smart.
Understanding the limits of what version 1’s path can deliver, and daring to think that pattern generation is more important than pattern recognition when it comes to delivering an imagined reality is just part of what’s so exciting about what’s currently going on behind closed doors.
And soon, those doors will open.
Toby Simpson, CTO and Co-founder, Fetch.ai