The assembly of algorithms

At the borders of ethnography, the history of computer science and the philosophy of science, Florian Jaton sheds light on algorithms, these black boxes which guide financial decisions, loves and sorrows.

For twenty years, algorithms have been presented in public debate as a power in themselves. Since the formula “code is law” launched in 2000 by the jurist Lawrence Lessig, many authors have argued that automated data processing governs practices, replacing law and politics. Recommendation systems built from large databases thus guide decisions and tastes on the Internet, in the areas of purchases, access to information and cultural consumption, governing romantic encounters on dedicated sites. , the application of sentences with predictive justice, or financial investments with decision support algorithms. Algorithms thus vectorize society through recommendations and arbitrations that are all the more powerful as their procedures turn out to be largely invisible. Florian Jaton proposes to shed light on these techniques with black box status. To do this, he takes a path at the borders of ethnography, the history of computer science and the philosophy of science. With the exhibition of the methods of producing an algorithm, it puts the history of computing into perspective in order to transform esoteric techniques into objects of controversy.

The laboratory animal algorithm

To keep this promise, the ethnographer turns away from the algorithmic blockbusters of Silicon Valley to immerse himself, like B. Latour and S. Woolgar, in the life of a laboratory. The author follows a small IT team from Lausanne, specialized in the field of automatic analysis of visual “saliency” (p. 59), a branch aiming to produce image processing systems based on the vision model. human. Through meetings, coffee breaks, blackboard sessions, hallway discussions, emails and programming sessions, he becomes a full-fledged member lending an unsteady hand, but helpful enough to observe the stages of developing projects in a situation. ‘an algorithm.

Supported by public funding, this laboratory is dedicated to the production of academic work based on close links with large digital companies and the world of startups, through invitations to presentations, collaborations, paid internships. , recruitment, and the network of conferences which validate and bring recognition to several months, sometimes several years of research (p. 40). After two and a half years of immersion, the author leaves with a thousand pages of handwritten notes, two thousand .txt files, dozens of scripts coded in Python, hundreds of images and video recordings (p. 45) . From this dive, he carried out a series of observations, which constituted a first breakthrough in a field where the functioning of machines and the architecture of programs cover the deepest mysteries for ordinary mortals. From the time spent scrutinizing unspectacular technical activities and sometimes decidedly boring tasks, the author draws a robust proposition of what an algorithm is, distancing himself from what is usually said about it.

Algorithm, did you say algorithm?

It is in fact common to define algorithms as a method intended to solve a problem using steps that are sufficiently unambiguous, finite and systematic to be followed. In its sum The Art of Computer ProgrammingDonald Knuth argues that algorithms aim to transform data into results, by transforming given quantities (“inputs”) into quantities of outputs (“ outputs “). The field investigation highlights the inadequacies of this conception. First, algorithmically processed data does not fall from the sky; they are the product of assembly work, which involves the use of tools marked by the desires, beliefs and constraints specific to the history of their designs. Furthermore, the acceptance that has become standard by D. Knuth minimizes “the infrastructure for evaluating algorithms and their political dimensions” (p. 49): the choices and orientations are not self-evident, but are the fruit of a series of decisions whose merits could be open to discussion. From this perspective, programming practices must be brought to the forefront, because they encapsulate a form of arbitrariness in the techniques.

The book is constructed as a recapture of this dark side, through the description of the production process. The three terms in the subtitle designate the stages. First, researchers create a coherent and algorithmically processable database. Second, they seek to make writing games coherent enough to effectively instruct computer programs, validated by communities of peers (p. 164). To do this, they carry out a series of reductions aimed at transforming non-mathematical entities (here images) into mathematizable entities (values, distributed in rows and columns). If mastered, this sequence of operations harmoniously links an ad hoc database (“ground truthing”) and programming activities (“programming”) based on mathematical equivalences (“formulating”). If the opening of this black box proves slow and delicate, it is because IT comes from a history marked by concealment strategies. The book attributes the founding act of this propensity towards ellipsis to John Von Neumann.

Computer science, between metaphor of the mind and the power of mathematics

In 1944, Von Neumann, one of the most influential scientists of his time, closely followed developments inElectronic Numerical Integrator And Computer (ENIAC), first fully electronic computer designed at the University of Pennsylvania. For the needs of the Manhattan Project in which he is directly involved, the mathematician wishes to develop a superior machine in terms of computing power. The war effort made it possible to open lines of credit to finance this computer, called Electronic Discrete Variable Automatic Computer (EDVAC). Its design pits three groups against each other: a team of engineers who attempt to translate ballistic data into electronic diagrams using differential equations; operators who take care of translating its diagrams into signals understandable by the machine using electrical wires and control buttons; and Von Neumann himself, inhabited by a vision of the computer relatively far from the concrete and often chaotic experience of the first two groups.

Attentive to the most recent scientific developments, Von Neumann found in neurology a model for the computer, which he envisaged as an “electronic brain” (p. 91). Two researchers have just proposed a simplified and mathematized representation of the human brain functioning like a network of neurons that communicates via electrical signals. This analogy proves decisive in the history of computer systems, often conceived as schematized and mathematized representations of the brain. Recalling the weight of cognitivists in this story, the author depicts computing as a continuous tree of decisions. Von Neumann’s initial orientation is due to his attachment to abstract models, which allow the identification of the human mind with machines that are as “boring as dishwashers”. This reduction opens the way to a history of computing based on the alignment between material entities and mathematical entities.

Indeed, the group studied by F. Jaton constantly consults, manipulates and matches and compares values. When entities do not present a mathematical form, they work to make them mathematizable, in order to make them symbols that are more easily “shareable” and “comparable” (p. 233). When they succeed, the boundaries between humans and machines are partly lifted, since it becomes possible to send computers a list of instructions to follow and repeat. Space constraints are then reduced: the author gives the example of the database used by the team, made up of anonymous and remote workers, paid by microtask on the principle of crowdsourcing. The time constraint is also reduced since the programs carry out the calculations in a few seconds where a lifetime would not have been enough for human work alone.

From the black box to a society of algorithmic controversy

The author was able to arm himself with references to carry out this patient and rigorous work of elucidation. Because if it is alone that he crossed the threshold of the laboratory, he mobilizes throughout the pages the community of Science and Technology Studies. The different generations of the Center for Sociology of Innovation feature prominently, with the work of B. Latour as the center of gravity. This tradition of research, fruitful, but sometimes arduous and a source of misunderstanding, is mobilized with great clarity, which makes it a beautiful introduction. in vivo. The author’s challenge to shed light on the procedures for developing an algorithm therefore seems honored; on a scientific level at least.

Because there remains the other side of the book’s promise, more directly political. However, if the author precisely reconstructs the way in which an algorithm is gradually built locally, it does not entirely allow us to understand how those of Amazon, Facebook, Google or Netflix proceed on an international scale to guide the behaviours. By focusing on a case study that can be described without offense as peripheral, the book almost involuntarily reports on a state of affairs that is rarely mentioned in public debate: in the majority, algorithms are appear to be far removed from any capacity for political influence. Limited in their application, little used, or even really recognized by communities of experts, they are the subject of harsh selection throughout life cycles that are sometimes very ephemeral, as experienced by the small team that sees his first proposal rejected by a renowned conference. Most are thus doomed to oblivion or relative anonymity, which essentially make them practically ineffective instruments of governmentality.

For others, whom the author calls for to be placed at the heart of public debate, the book ultimately says nothing about the conditions, means and forms of democratic criticism. If Bruno Latour proposed a Senate dedicated to environmental issues and Michel Serres, a world parliament composed of scientists responsible for the causes of life, it remains to be defined what an assembly of algorithms would be: what modalities (electoral, representative, delegation, co-optation, nomination , etc.), what risks to avoid (populism, technicalism, clientelism, corruption, etc.), what procedures for putting them on the agenda? The author must therefore at a minimum to readers a volume 2, especially since if the constitutional texts do not represent a summit of general public literature, the source codes do not yet play the role of national daily newspapers in the public debate. For all those who are interested, F. Jaton has just delivered a true learning novel.