What the coming year holds
While progress in the development in AI is now rapid, the implications for jobs and human intelligence are misunderstood, and the scope for human intelligence is enhanced not diminished - at least in 2018!
AI and the future of work
There is a revolution underway driven by deep learning. After multiple false dawns, artificial intelligence is extending human capability generally and business performance specifically. This has led to predictions of a ‘hollowing-out’ of human experience; computers will go beyond taking away the drudgery of repetitive analytics and supplant humans in the very activities that we think define us. But this is to misunderstand the strengths and weaknesses of this new technology.
Powerful and impressive as artificial learning is, AIs are still very narrow in their field of operation compared to human intelligence. Deep learning networks are good at interpolating between examples, or extrapolating a short way from them. But they do not have the general experience or structural flexibility to recognise and solve new problems without training. So in the medium term, human intelligence will not be confined to producing training samples for cleverer machines, or to predominantly interpersonal roles as organic ‘front of house’ staff: wherever there is a new problem, more general intelligence will be required - just amplified and made more potent with AI assistance.
AI and jobs in January 2018
The deep learning revolution has allowed computers to execute tasks which are hard to codify, but easily repeated by (and repetitive for) skilled humans - email spam identification, image classification, voice recognition. This has immediate practical application, and is empowering of human roles rather than undermining them.
One of the uses we turn AI to at Filtered is the tagging of training content - associating training articles, podcasts, videos etc with tags that make them easier to search, filter, prioritise and use.
For a neural network to perform this tagging task, we need to teach the network which tags should apply to which content by example. We build a ‘training set’ of examples of well-tagged assets. Expert human input is critical at this stage, as it defines ‘good practice’ for the algorithm.
Through an iterative process, the neural net learns which tags should fire in response to which inputs - just as synapses in the brain are set to fire in response to familiar inputs. Once the training is complete, we have a neural network that can associate new assets with identifying tags. Human expert decision making will be amplified by the algorithm because it can apply the tags to huge volumes of content.
Humans are still the best at tagging most new data: manual curators can categorise content quickly and in good agreement with each other. But it is time consuming and costly. AI does the same work at scale, by embodying human decision making in a trained neural network.