| Not So Hard Any More |
| Written by Mike James | ||| | Wednesday, 29 July 2026 | ||| | I was looking over some of my favourite xkcd comics and one made me think more deeply about what AI has done to us. It started from a realization that AI had spoiled a good cartoon to something a bit deeper - we never intended this. I don't imagine there were many AI researchers back in the 1980s who thought that the road would lead to this ... but first the comic: It's a good joke or it was. We programmers could laugh at the ignorance of the non-CS educated. Of course, it's easy to work out if someone is in a national park - just download the GPS and then use a map, GIS, to figure out the answer. A simple lookup, once you get beyond the hardware needed. But then there is question of whether the photo is of a bird or not - yes very difficult because, at the very least, it needs some computer vision and perhaps some AI so yes it is virtually impossible and a team and five years is hardly enough. This comic was published towards the end of 2014 and it indicates how innocent we were. There was already progress on convolutional neural networks in 1989 to 1998 with LeNet able to read simple hand-written text. In 2014 we had GoogleLeNet, which was certainly up to the task that was proposed to need a five-year research project. So even in the year that the comic was published, the problem wasn't so very difficult and just a year or two later the performance of such networks was well up to the job - making it almost too easy. The key phrase that struck home for me was "the easy and the virtually impossible" by this standard we are into the realm of the virtually impossible. Of course, the solution has some problems. Our current crop of neural network-based systems all share a feature - a flaw perhaps. Convolutional neural networks are vulnerable to classifying things that look nothing like the target. We all recognize that LLMs are prone to halucination and it's the same phenomenon. Networks of this size are not fully deterministic in the sense that small changes in the inputs can produce crazily different outputs that really don't correspond to reality at all. So imagine our bird photo detector, working well apart from a few photos of birds that it classifies as something quite different, and by the same token some things that are not birds as birds. We have swapped virtually impossible for a solution that isn't provably correct in any circumstance. For some things, this isn't too important - who cares if a few bird photos are missclassified. In other cases, it can be very very important. Everyone cares if the vision system of a self-driving car classifies a person as something else. I said as the start that we never intended this and it is true. Back in the early days of AI it was all an engineering ethos. Computer vision, for example, was deterministic and the filters and tools were clearly defined and while they might not give the correct answer on difficult targets, we understood what they did and it was assumed to be a matter of time before we fine-tuned the approach to make it as close to 100% right as we needed. Even the nascent neural networks of the time were thought to be trainable to produce deterministic outputs - any idea that thy could hallucinate or mis-identify things in a mysterious way was not even suspected. We didn't think that adopting AI would mean un-adopting hard engineering practice. I am fairly sure that back then no-one imagined that LLMs would be used to produce code that was potentially buggy and insecure and even more unimaginable was the idea that if they did anyone would make use of it! ## Related Articles To be informed about new articles on I Programmer, sign up for our
Comments #
or email your comment to: | ||| | Last Updated ( Wednesday, 29 July 2026 ) |