• With no need for an IT professional to design and build new applications, systems developed using generative AI will be more likely to fit the specific needs of their users, increasing the probable effectiveness of the applications
• Nonetheless, many IT staffers with whom we’ve spoken are opposed to citizen development. They fear that it will result in poor-quality systems that the IT team will have to repair—or that generative AI tools will replace IT entirely.
• How will they facilitate and safeguard citizen development without putting too many obstacles in its way? To reject its benefits is impractical, but to manage it carelessly may be worse
• Facing a lack of professional data-science talent in many of the countries in which it operates, ING began to explore citizen data-science capabilities
• For example, what happens to a citizen-developed application if a department comes to depend on it, and the employee who developed it leaves for another company?
• We do know of some examples of citizen-developed programs that briefly went astray.
• RPA can be misused to automate one’s own tasks in order to take on an additional job in another company. Executives who have voiced concern about this envision workers secretly holding down two or three full-time jobs with the help of RPA or generative AI, and we’ve found online discussions in which such “overemployed” individuals describe their experiences
• PwC has established a broad process for citizen development that it calls the “digital accelerators” program. Launched in 2017, the program originally had three components: data science, automation, and data management. It was recently revised to include low-code/no-code solutions and productivity improvements. The company has announced plans to teach its 65,000 employees AI skills, including generative AI.

6 more annotations...

• Prior research has demonstrated that decision-makers are often reluctant to accept AI-generated guidance, whether consciously or unconsciously, and routinely overrule it.
• Turns out, shipments were up to 50% closer to the recommendations generated by the black box system than to those from the simpler, rule-based algorithm, suggesting that employees trusted the black box model much more.
• ur finding runs counter to conventional thinking, which holds that people are more likely to accept guidance from systems they can understand.
• decision-makers believe they understand a model’s reasoning better than they actually do.
• they knew the model had been developed and tested with the input of some of their colleagues, and that gave them confidence in it
• It’s not like we trusted each recommendation the model made. It’s that we trusted the model on a more macro level, because our peers had been involved in its development
• Stocking decisions for high-volume stores evoked the most uncertainty because the consequences of misallocations were greater: potentially higher amounts of lost sales or excess inventory. People were far more likely to accept the uninterpretable model’s recommendations than those of the interpretable model when allocating products to those stores
• You’ve got to constantly measure how much employees trust your system. If you find that they trust it in some contexts more than others, you’ll want to dig into why. They might be skeptical of the AI because it’s buggy or is offering recommendations that are vastly different from what their experience or intuition tells them.

6 more annotations...

• En effet, si l’on attribuait une goutte d’eau à chaque calcul nécessaire à la formation du modèle PaLM de Google, par exemple, le volume ainsi constitué suffirait à remplir l’océan Pacifiqu
• Microsoft a signé, dans le cadre de son objectif de 100% de consommation d’énergie sans carbone d’ici 2030, le plus important contrat d’achat d’énergie renouvelable jamais conclu par une entreprise, pour un coût estimé entre 11,5 et 17 milliards de dollars
• Si l’entraînement d’un grand modèle de langage coûte 126 000 litres d’eau, une conversation avec ChatGPT en dépense environ un demi-litre.
• Amazon Web Services va payer jusqu’à 650 millions de dollars pour installer un centre de données en Pennsylvanie à proximité d’une centrale nucléaire
• utiliser ChatGPT pour résumer un email revient à se faire livrer une pizza par une Lamborghini
• Or les grands modèles de langage ont jusqu’à présent progressé grâce à l’augmentation du volume de données sur lequel ils étaient formés. Si les grands modèles continuent de dominer aux dépens des modèles plus ciblés (cf. supra), le risque est donc qu’ils soient bientôt partiellement formés avec des données synthétiques, c’est-à-dire des données créées par des intelligences artificielles qui perpétueront biais, erreurs et désinformations.

4 more annotations...

• we’ve been using AI but more in technical ways of [automating] customer service and improving our logistic positions.
• “In our line of business, people search for products, finding and buying them. [AI] is about being accessible and making [that process] easy. People are looking for inspiration
• It used to cost $10,000 to translate one 10-hour course into one language. We can now do that for $20.
• [AI] is moving us forward to a future where [the solutions to] more problems that couldn’t be solved today are starting to become more accessible—you’re on some sort of intelligence continuum

2 more annotations...