One of the three films with the most Oscar nominations, it wins the statuettes for Best Film, Best Direction and Best Screenplay, among other awards. In a sublime final scene, the screenplay hints at a coming change in the audience’s preferences: many of the dangers that threatened Margo at the beginning of the film now await Eva.
Digitalisation, automation and robotisation of processes in relation to artificial intelligence
We have been hearing about digitisation, automation and robotisation of processes for years. Recently, the emergence of Artificial Intelligence (AI) has relegated the former to the background, making its way into our everyday conversations with the veiled promise of solving our needs in an immediate and simple way: “ask the chatbot”. Driven by such an effective narrative, more and more organisations with no prior experience in this area feel the compelling need to jump on the AI bandwagon.
Such apparent simplicity is counter-intuitive at first glance. As defined by Stuart Russell and Peter Norvig in their excellent ‘Artificial Intelligence: A Modern Approach‘, Artificial Intelligence is the field of knowledge that seeks to design solutions, whether algorithms, robots or machines “… that can act efficiently and safely in a wide range of novel situations”. The promise of automatic adaptation to unexpected situations is precisely the distinguishing feature of AI compared to its predecessors such as automation or process robotisation, which are also committed to eliminating human intervention, but require direct intervention via re-engineering to incorporate novel scenarios.
In increasingly sensorised environments, Artificial Intelligence takes on the role of Eva in Mankiewicz’s film, while the automation and robotisation of processes, that of Margo Channing, her mentor: many of the dangers faced by the veteran diva will end up being inherited by her young and talented successor.
One of them is the danger of working on certain stages. Despite the advances made by Artificial Intelligence in recent years, neither machine learning nor, of course, its predecessors in viewer preferences guarantee a zero error rate. And this leads us first of all to reflect on the consequences of these depending on the scenario where the performance of our new star takes place.
Imagine an AI participating in a chess tournament. Any possible mistake made by the AI, however embarrassing, does not endanger the integrity of the opponent or expose the team leading the development and training of its models to possible legal action.
Now let’s think about autonomous driving in a car: it’s a different story, isn’t it? In such a peculiar scenario, classical automation also faced similar risks: this is why hybrid or supervised solutions are proliferating in sectors such as the automotive industry, where sensorisation is becoming increasingly important. Imagine, for example, parking in a state-of-the-art car; if the sensors detect the proximity of an obstacle, the vehicle’s software will alert us to the potential danger. However, it is the driver who will have the final say.
We called the simplicity promised by AI counter-intuitive, as the training models on which it is based require a huge effort on the part of suitably qualified human personnel.
Today’s sensorised world helps to accumulate fabulous amounts of data, so two tasks become particularly important: data cleansing and data categorisation.
Data cleansing and categorisation
The need for these two activities can be better understood with the help of another simile: the incorporation of video refereeing in decision-making during a football match. In particularly controversial plays, the referee can momentarily interrupt the match to review with the help of VAR the available images, taken from different angles, until he chooses the most appropriate one. It should be noted that the camera with the best possible orientation does not always help to clarify doubts if, for example, another player comes between it and the protagonists of the action. In the world of Artificial Intelligence, the task of discarding shots would be part of data cleaning.
Even with the help of video refereeing, the decisions that the referee ends up making can be very controversial, which is why they periodically meet to review plays pre-selected by a committee with the aim of establishing common criteria. Some of them will even be part of the syllabus used to train new generations of referees. Returning to Artificial Intelligence, this is the role that categorisation takes on, since, based on a collection of ambiguous but documented situations, a decision criterion is established with which to train (and retrain) the model for future decision-making.
From the end consumer’s perspective, the promise of simplicity that accompanies AI is correct, because as a viewer they do not perceive this layer of complexity: the work of cleaning data and categorising it is a behind-the-scenes activity. But while the Margo Channing’s of our modern history require re-engineering, it is common in the most advanced AI models to rely not only on initial training (the letter P in the acronym Chat GPT stands for ‘pre-training’) but on successive re-training, through fine-tuning techniques in which the involvement of human specialists is still necessary.
Other challenges of artificial intelligence
We end this quick review with another challenge that AI shares not only with its predecessors (automation and robotisation of processes) but with any project. We are referring to the difficulty we humans have in defining our own needs and prioritising them, as George H. Gallup anticipated almost a hundred years ago when he laid the foundations of market analysis. Who has not suffered in their own flesh some colossal misunderstanding resulting from the process of capturing requirements, or in the interpretation of these?
But this is another story that deserves to be told in more detail, and we will leave it here for today, because as Margo Channing said: “Fasten your seatbelts, it’s going to be a bumpy night!“.
