Most so-called AI software is based on machine learning.
When comparing conventional software with machine learning software, note the following: To make conventional software, software developers must write out each step to create a specific software product. On the other hand, to make machine learning software, software developers must collect relevant data. The data is used to “train” the software. After this training, the software is available for use. No step-by-step programming is needed.
Both types of software require effort to produce, but machine learning software requires developers with a different skill set than what is needed for conventional software. Neither type of software is perfect. Therefore, both types of software must be tested to look for problems.
Each example of machine learning software is based on a model, and models vary in size. Current LLM (large language models) have over a trillion parameters. However smaller models exist.
It is in theory possible to construct conventional software that can be exhaustively tested. Machine learning software with large models cannot, even in theory, be exhaustively tested.
Machine learning is not software that “learns” in the normal sense of the word. Software can be written so it changes its behavior over time in response to input, however that is not necessarily true of machine learning software. And in any case, no software is as good at learning as a typical human.
There are many potential applications for machine learning, some have been more successful than others. A few examples:
In 1950, Alan Turing made an estimate of the computational capacity of the human brain and concluded that a computing machine that duplicates the human brain would be possible in a few years [9].
Turing was wrong; 75 years later we are not any closer to this goal. To be sure we’ve built computers that play chess, computers that seem to understand human speech (such as Siri and Alexa) among many other things. But imitation is not duplication.
We have learned a lot since Turing’s time, primarily by studying animal models. For example, studies using domestic cats have led to insights into the visual system of mammals [10].
See Norden, 2007 and Haier, 2013 for a discussion of the human brain [11,12].
We’ve also learned a lot by studying simple organisms. Caenorhabditis Elegans (C. Elegans for short) is a nematode worm about one millimeter in length. An adult worm has a nervous system with 302 neurons. It has been extensively studied, which has resulted in a complete wiring diagram. In spite of all that, said nervous system is not completely understood [13]. If we don’t understand a nervous system with 302 neurons, we definitely don’t understand a human brain with 90 billion neurons.
For a discussion on how intelligence has evolved over time [14].
Predictions about the future of technology are frequently wrong [15]. The future of machine learning software is no exception. Typically, predictions about machine learning take the form of a narrative that goes like this: machine learning software is evolving rapidly and will continue to evolve, producing more and more “intelligent” software. However, the narrative is simply wrong.
Machine learning software depends on data; to improve the software you must improve the data used in training. In the early stages, adding more data to the training set improves the software. However, after a point, adding more data does not improve the software. The point at which things do not improve is known as “peak data”[16,17,18,19].
Existing large language models (for example chatbots such as ChatGPT) have reached peak data, and no significant improvement is possible. It is possible to add bells and whistles. It is also possible to make improvements; but if improvements are made in one area, things will get worse in another area.
While the goal of making software “more intelligent” (however you wish to define that) is not realistic [20,21,22,23,24,25], there is a path forward. It is possible to build machine learning software based on small models. Unlike large models which attempt to do everything, small models are limited in what they can do; they don’t try to do everything. They aren’t intelligent. They are less expensive to produce, require less data, require less compute time to train than large models. They don’t require large numbers of data centers. However, within limited domains they can be very useful.
The emphasis on machine learning over recent years is likely to have a counterintuitive effect. This emphasis has prompted many businesses to think about using software to improve efficiency. In some cases, the answer will be machine learning software, but in other cases the answer will be conventional software. It is likely this will lead to an increased use of conventional software in business but might not lead to a significant increase in use of machine learning software.
For general information about machine learning, see Géron, 2022; Littman, 2020; Grim, 2026.
For general information about the problems and limitations associated with machine learning software, see Dignum, 2026; AIAAIC.org, undated; Olavsrud, 2024, Omwanda, Emman. 2024. Lewis and Mitchell, 2025.
Specific notes:
See Book List for AI and Machine Learning