Will generative AI lead to the death of software?

Pléiade AM's Orientation and Foresight Committee brings together a group of technology professionals working in the fields of AI, cybersecurity, data management and business applications. They meet regularly to discuss technological developments and help the Pléiade AM team understand the technological and market challenges facing Cloud companies.

It's an abrasive question, but the emergence of generative AI in recent months is akin to a kind of Copernican revolution in digital technology, which has been dubbed the "iPhone moment" to inadequately capture the changes to be expected from the spread of this technology. Today, we know that generative AI, based on large language models (LLM), will lead to upheavals in economic, creative, legal and even medical processes.

Much has been said about the risks to employment in certain professions, and the staggering productivity gains to be expected in the years to come. But many questions remain unanswered about the changes to be expected from generative AI, and it will be difficult to claim definitive answers at this stage.

What will be the value chain of generative AI: will it favor financially-armed tech giants, or new players still in limbo?

Of all the questions that generative AI has raised since it first broke into our daily lives, one caught the interest of Pléiade AM's Comité d'orientation et de prospective: why should software, which was the linchpin of the wave of corporate digitalization that preceded the emergence of generative AI, be particularly threatened by it? And first of all, what will be the value chain of generative AI: will it favor the financially armed tech giants, or new players still in limbo? These are the questions that drive Pléiade AM's Comité d'Orientation et de Prospective, on which global tech professionals sit, in meetings that seek to bring out avenues of reflection on these subjects. Here's the current state of play.

The generative AI value chain

When you consider the sums poured into Nvidia for its precious GPUs, and the cascade of funds raised for generative AI start-ups ($10 billion from Microsoft for OpenAI, $4 billion from Amazon for Anthropic and €100 million for the French newcomer MistralAI), you have to wonder how all this is going to pay off, and who the winners of generative AI will be.

Shovel salesmen during the gold rush didn't care how happy prospectors were.

Today, it's clearly the Nvidas and AMDs of this world, i.e. the suppliers of the tool that trains LLMs. At this point, the cost of generative AI is obvious: shovel salesmen during the gold rush didn't care about prospectors' bliss. ChatGPT's success was instantaneous, first among the general public, but also through the diversity of new feature announcements by the software's players. As a result, the subject drew the attention of OpenAI's competitors, who in turn released LLMs (Meta, Google, Anthropic), and then the open source community.

At the same time, Moore's Law, combined with new, less parameter-intensive training methods, has enabled the open source community to seize on existing LLMs and further reduce training costs. So-called "foundational" models, such as OpenAI's ChatGPT or Anthropic's Claude, continue to be expensive to train, as they aim to improve their performance by increasing the quantity of parameters. They can then be refined at lower cost, with data sets specific to an industry or even a company, in order to provide answers more suited to a particular type of activity. These are vertical models.

It's the near-universal spread of generative AI that will drive the value chain.

In the end, it's the near-universal spread of generative AI that will drive the value chain: chip manufacturers and designers of foundation models will certainly benefit. But it is probably end-users who will reap the greatest benefits, through the massive productivity gains they will achieve, notably through the burgeoning functionalities of enterprise software.

But if software is the pipe through which generative AI spreads, why think that it is in danger of obsolescence by the very fact that generative AI is spreading?

Why software has invaded the enterprise

The softwareization wave of recent decades has led companies to "standardize" their procedures.

What is software? Software is a set of standardized, digitized procedures with a user interface. It enables data flows to be interconnected, thereby automating or optimizing certain business processes. The wave of software development over the last few decades has led companies to "standardize" their procedures to fit into the boxes of the various applications used.

Who would have thought that laundromats could be "softwareized"?

The deployment of the Cloud from the 2010s onwards not only facilitated access to software, with the SaaS (Software as a Service) wave, but also enabled software to conquer new uses by replacing physical processes. These include dematerialized signature software (Docusign), or medical workflow management (Doctolib). But the process is still in its infancy. Who would have thought that laundromats could be "softwareized"? Yet this is the service provided by American company Cents, which connects washing machines for use by customer applications, and automates the management of consumables and financial flows for the manager. This company already generates over $100m in annual recurring revenue from 2% of laundries in the United States. Thanks to the Cloud, we're entering the golden age of software.

So why should we think that generative AI could threaten this deployment, which is still in its infancy? All the more so as the 1st steps of generative AI in our world have instead resulted in a blossoming of functionalities in these same software offerings, making them even more efficient and therefore more attractive.

There are 2 elements that may call into question the foundations of software, or at least of part of the industry: the automation of IT development, and the domination of data over the software interface.

The automation of IT development and its possible consequences

The LLM, launched into a dataset of billions of parameters, has learned to code on its own.

ChatGPT's training led to a major development that was unexpected by its designers: the LLM, launched into a dataset of several billion parameters, learned to code by itself. This new know-how was promptly put to use by the developer community to write chunks of code, thus considerably boosting developer productivity. But it's in the nature of LLMs to constantly improve, so we can legitimately wonder about future needs for human code writing.

From today's point of view, whether we consider Copilot (GitHub) or Code Whisperer (GitLab), depending on the development platform used, productivity gains for developers should rapidly reach 50%. Such development can lead to 3 possible consequences:

• Consequence 1: Such a revolution means that more products can be produced, more quickly, and at lower cost. It therefore supports the idea that generative AI will, instead of threatening software, enable it to insinuate itself, thanks to the Cloud, into all areas of the economy.
• Consequence 2: We may well wonder whether the drastic reduction in production costs might not lead to the emergence of new players likely to threaten established companies, with lower development costs and therefore lower sales prices.
• Consequence 3: We can even imagine, in the not-too-distant future, code being completely taken over by the machine, and software functionality being created by a simple "prompt" in natural language. We can therefore easily imagine that this democratization of development could lead industrial or service companies to produce the specific functionalities they need themselves, rather than bending their procedures to fit the format of "shared" software. It's interesting to note that 76% of Fortune 500 companies are customers of GitHub, Microsoft's code management platform, which includes Coca-Cola and Fedex, not exactly technology companies. But what company won't be a "technology company" in the future?

Data dominates the software interface

Generative AI is the most advanced tool for "making data talk".

The Cloud has enabled the exponential growth of data, but above all its collection and organization for exploitation. New tools that organize data in data lakes, data warehouses and databases enable companies to manipulate large quantities of data, and transform them into actionable signals to generate productivity gains. It's easy to see why generative AI is the most advanced tool for making data "talk". Today, an LLM is capable of ingesting data and rendering it in the form of an image, table or text, with almost no limits on form. In other words, once data has been collected, even in very raw form, it would be enough to apply a generative AI model to it to obtain any structured information, be it accounting, financial, sales or marketing. In such a context, would it still be necessary to buy a rigid software interface, such as a CRM or ERP? Could a simple "prompting" interface replace the thousands of screens, menus and click boxes of our business software?

We are therefore entering the data era, after the software era. It would certainly be hasty to predict the death of software at this stage. Perhaps it will even increase its presence in our lives, thanks to artificial intelligence. But there is some uncertainty about its long-term survival in its current form, as a man/machine interface in business applications. On the other hand, the new data management tools that have emerged with the Cloud will be central to companies' ability to extract all the productivity gains and make the most of all the technical possibilities offered by generative AI.

History continues to be written, and we will be vigilant observers.