Programming in natural language


Jaak Henno, Tallinn University of Technology, Estonia
Hannu Jaakkola, Tampere University, Pori, Finland
Jukka Mäkelä, University of Lapland, Finland

This presentation is on-line at http://staff.ttu.ee/~jaak.henno/conf2025/index_ext_min.htm

Data, the most important modern raw resource, grows rapidly

more than 90% of this data remains unused


IT investments are growing


IT projects are failing

33% of the global software development money is written off annually
IT has become a BIG WASTE OF MONEY

The main reasons for IT misery: :

COMMUNICATION - Humans and computers do not undersand each other languages

Programming topics have become very complex

1975: Create a program to calculate sum of integers from 1 to 100...
2025: Create a program to calculate the amino acid sequence coded for by the DNA sequence: 3' T A C C C T A G A 5'


All problems are first stated in human language


Programmers have to translate them to computer language

Humans use 30000-35000 words
To English language is annually added 1000 new words

Programming languages use less than 100 symbols
Translation of a task from human language to computer language creates structure with enormous depth

Description of the task (in English) to create a Python program to draw this chart had 34 words

The python program contained 36 lines

The program loads 1175699 lines of code (LOC) from libraries

Nobody knows exactly what these 1175699 lines do

For processing (big) DATA are increasingly used programs based on Large Language Models (LLM) - chatGPT, Gemini, Perplexity

These programs can produce fluent text on several natural languages (English, Estonian, Finnish, Croatian....)

Although these 'fluent chatterers' do not understand a word what they output - they are also able to create (simple) computer programs

We tested the programming ability of several big LLMs (chatGPT, Gemini, Perplexity etc) : :

1. To LLM was given the task description, e.g.

Create a Python program illustrating software project statistics...

LLM outputs a Python program, which we execute

LLMs can repeat programs and use libraries which were stored in its training Corpus.

We investigated LLMs abilities with the 'Hello World' example of the Data Science - the iris data set introduced in 1936 by the British statistician and biologist Ronald Fisher

The iris data set is a mapping from 4 real numbers (the length and the width of the sepals and petals, in centimeters) to three names of iris species (Iris setosa, Iris virginica, Iris versicolor)
The iris database has been considered in many papers; the main method has been considering 4 real numbers describing a plant as a 4-dimensional point and investigate distanes between these points, trying to find clusters of closely related plants, i.e. compress information.
We give similar task to chatGPT, replacing species names with color:

Import list of items from the uploaded file iris_col_dat.py, calculate distances between list items using the first four elements of lists, create graph of items connecting items with minimal distance, color the graph nodes using the last element of the item as an RGB-triple

Items from the same Fisher's cluster appered as 42 connected components, but this number is too big to consider it as information compression

The plant attributes - sepal-petal width-height - have strong dependencies - they grow when plant grows, but maintain similarity, what is seen when width-height of Iris Setosa samples where used as width-height of concentric rhombuses:

Thus plant data should be considered not as 4-dimensional points, but as 4-dimensional vectors. Vectors similarity appears as minuscule angles between them.

Replacing the model of data from 4-dimensional points to 4-dimensional vectors and considering angles instead of distances is a step, what LLMs can not do, here is needed human intelligence

We always study the WORLD in some finite model of some finite fragment of the World, e.g. study iris flowers as a mapping from 4-dimensional data points to names of species

Considering 4-dimensional iris data not as 4-dimensional points, but as 4-dimensional vectors is a change of iris model.

This is something what LLMs can never do

But when we (humans) change the model, we can use LLMs to investigate this new mode

To get overview we asked chatGPT to create histogram of angles between items and a fixed vector - the average of all vectors.

"Create a Python 3 program, which imports from the file iris_col_dat.py the list items, finds the mean of the list and creates a histogram with 24 bins of angles of lists of first four components of lists with the mean; set color of 3 locally maximal bins of the histogram to 'aqua'."

Expectation of items in three local maximums were added to the list of items as new 'centers of attraction'

Create a Python 3 program, which imports from the file iris_col_dat.py the list of lists Items, using the first 4 items as coordinates finds for each list except the last three distances of the list with the last three lists, finds the minimal distance among these three distances, creates a graph whose nodes are lists from the list Items; for each node except the last thee add edge to node in last three nodes which had minimal distance with the node; color nodes using the last item in the list as the RGB triple for node color

By popular clustering criteria (the Davis-Bouldin-Index, Silhouette score, Calinski-Harabasz Index) this clustering is better than Fisher's

Thank You !