Why neural networks cannot replace developers yet

rabiarahi

The excitement around neural networks is not abating. AI algorithms trade on stock exchanges, drive cars, help investigate crimes and compose music in the style of Yegor Letov. In general, we say hello to Will Smith with the film "I, Robot" and begin to politely greet smart technology, counting on mercy after the uprising of the machines.

Things are moving towards the fact that artificial intelligence will soon replace representatives of any profession, not sparing even sugar daddies-programmers. Should we be afraid of this? How quickly will neural networks be pumped up to at least the level of juniors? Will the struggle for work and survival begin? Let's figure it out together with the mentors of the IT school TeachMeSkills!


Copilot and ChatGPT are the most likely "replacements" for programmers

The capabilities of AI are limited by two things - computing social media marketing service
power for their training and time. You can teach a computer to analyze photos and distinguish Monet from Manet in about 20 minutes. The most hyped, daring and interesting to most programmers "neural programs" are ChatGPT and Copilot. The first is just a household word, and the second was created specifically to make developers' work easier:

The Copilot neural network fully lives up to its name — it is a second pilot, a partner, and a virtual colleague. The smart add-on is capable of analyzing the code and offering options for its continuation in the form of separate fragments and entire functions.

ChatGPT, on the other hand, acts as a general-purpose neural network. It answers user questions using a large database. Helping develop computer systems at the level of a novice programmer is just one of the chatbot's options.
The Copilot neural network fully lives up to its name — it is a second pilot, a partner, and a virtual colleague. The smart add-on is capable of analyzing code and offering options for its continuation in the form of individual fragments and entire functions.



ChatGPT, on the contrary, plays the role of a general-purpose neural network. It answers user questions using an extensive database. Assistance in developing computer systems at the level of a novice programmer is just one of the chatbot options.
"Globally, we are talking about two intelligent systems, tailored for different tasks. If Copilot is strictly for programming and writing code, then ChatGPT is about everything at once," says Maxim Stepanovich, mentor of the Machine Learning course at TMS.
Three Problems with Neural Networks

Indian tech support company Duukan fired 90% of its employees and replaced them with artificial intelligence. American publication BuzzFeed laid off 180 journalists because of ML algorithms. Perm-based Xsolla team got rid of 150 specialists because the neural network considered them "low-productivity".

The statistics are alarming, but they do not concern programmers. There are at least three reasons why artificial intelligence algorithms do not allow developers to be driven out of their professional nests. Today, AI acts only as a tool to help cope with intellectual routine.

1. Weak understanding of context

The neural network does not understand the context of the project for which it writes code, despite all the efforts of the creator of the prompt (a smart request - a set of instructions used to generate the desired result). The machine does not understand the essence of the global task and does not understand the details of the implementation of the modules included in the product. If the specialist manages to get something working as a result, the code is most likely heavily buggy, terribly vulnerable, and poorly optimized.
"A developer doesn't just write code — he invents it, implements it, optimizes it, uses shortcuts, and so on. If you take a simple task — for example, making an elementary script or writing an HTML page — a neural network will cope with it, but only after a detailed explanation of the conditions. No tools — ChatGPT, Copilot, or their analogues — will help a person who does not understand the essence of the task," — Maxim Stepanovich, mentor of the Machine Learning course at TMS.

2. Legislation and liability

Liability for harm caused by artificial intelligence is a deep, delicate and controversial topic:

How to determine the degree of guilt of an autopilot that crashed a car into a truck at a speed of over 120 km/h?
Can a developer of medical software be judged for suddenly sending a person for surgery based on an incorrect diagnosis?
Who should investors deal with if they lose money due to malfunctioning trading bots?

Until lawyers and ethicists find answers to these questions, we can forget about the widespread introduction of neural networks into various professions. Therefore, machine learning algorithms remain tools - for decoding calls, finding errors in code and automating routine. Over time, everything may change, but certainly not in the next 5-10 years.