The new artificial intelligence tools, with ChatGPT at the forefront, will revolutionize the creation of text, images, videos or music. His ability to write code has had less impact, but it's just as incredible. As in other disciplines, in recent months there have also been apocalyptic articles written by experts entitled "The end of programming" or "ChatGPT will replace programmers in ten years". The key, however, is in the details: what exactly is it capable of programming?
First, it is able to solve challenges that deserve a lot of impact. The journal Nature published on Wednesday an article by researchers from DeepMind (Google) where they speed up the effectiveness of a C++ language algorithm that had been in use for decades by 70%: "We demonstrate how artificial intelligence can go beyond the current state of the art by discovering routines hitherto unknown," says the article. "What is really interesting is that the system learned to write a series of instructions without anyone telling it how," explains Armando Solar-Lezama, a researcher at MIT (USA) who had access to the article before its publication.
"They told him what instructions he could use and he started testing combinations. The only thing they told him was whether they were right or wrong. Only with that the system could infer what were the instructions that should be written and connected to generate the program, "adds the expert on AlphaDev, the program created by DeepMind.
Google merges its powerful artificial intelligence laboratories in full commercial race with the competition
These advances prove that artificial intelligence (AI) overcomes complex challenges. But how far can it go without human intervention? The Nature article explains that they posed the problem as a game for the machine, which had been specifically trained, and humans also guided the process. It is important, but not an absolute revolution. "People are very excited," Solar-Lezama continues. "With these AI models, we've seen that they can suddenly do things that were previously considered very difficult. But one question we have now is to see what they can do well and what they can't. Based on that, how can we reimagine programming to make it more effective? There are already tools that are starting to help programmers, but only as an assistant."
It's like a smart learner
That word "assistant" or "helper" is often repeated in conversations with computer programmers and teachers, as if AI were a savvy learner. Although the novelty is shocking, few foresee an earthquake in the trade. As explained by Brigitte Pientka, a professor at McGill University in Montreal (Canada) and co-author of a paper showing that ChatGPT would score a mark on an introductory programming course: "ChatGPT can be surprisingly and deceptively good at generating short programs. But it's not very reliable for now. I would say that in the future it will be more important to be able to evaluate and validate automatically generated programs to ensure a level of security and reliability." Programmers, therefore, will continue to be just as necessary, perhaps with a different role, according to Pientka: "Saying that we no longer need computer scientists thanks to ChatGPT is like saying that thanks to calculators and Excel we do not need mathematicians or statisticians."
"Saying that we no longer need computer scientists thanks to ChatGPT is like saying that thanks to calculators we do not need mathematicians"
Brigitte Pientka, McGill University
The same specialized magazine that published in January on the "end of programming", published this weekend an article entitled "AI does not help programmers", by the famous professor Bertrand Meyer. As a programmer, Meyer writes, you'd do well with an assistant that would keep you at bay and alert when you swerve. He continues: "But that's not what I get [from an AI]. I get the equivalent of a cocky, intelligent, well-read graduate, also polite and quick to apologize, but completely, and hopelessly, careless and unreliable. That supposed help is of little use to me."
In forums of programmers like Hacker News there are dozens of open debates about how and how much ChatGPT and other more specific applications help. The variety of responses is enormous, but the irremediable enthusiasm of a few months ago has waned. Solar-Lezama puts it in the context of another stage in the history of programming: "It will change programming as it has already changed in the last 20 years. And in the previous 20. Today is different than when I graduated. New programming languages also arrived that allowed you to automate things that people made by hand. It was easy to reuse the code someone had written. Today I can write extremely complex Python programs in 10 minutes that a generation ago would have taken months of a large team."
As society becomes more digital, it's critical to improve the code powering the world's computing.
Today in @Nature, we present AlphaDev, an AI system using reinforcement learning to discover enhanced computer science algorithms.
How does it work? 🧵 https://t.co/2ukGaPey5w pic.twitter.com/dSedrv7eik
— Google DeepMind (@DeepMind) June 7, 2023
Doesn't save that much time
AI has become that assistant that allows you to reduce somewhat the time spent on programming. But not in extraordinary quantities: "At first I was surprised to find that ChatGPT gave good software code," says Daniel Lemire, a professor at the University of Quebec (Canada), who does advanced programming work. "Copilot [a Microsoft tool] can write entire sections of code, as if it can read my mind. But sometimes he's wrong. It is difficult to measure the time I save maybe 10%. Writing code is a small fraction of programming, which includes design, testing, benchmarking, discussions, specifications."
"It's going to change programming as it has changed in the last 20 years"
Armando Solar-Lezama, MIT
For these reasons, the work of programmer seems at the moment a reasonable bet for the future. Despite advances in software automation so far, programmer employment is growing. The only certainty is that in the future there will be more code, largely thanks to the agility that AI allows, but not less programmers. "I predict that we will continue to hire more people, regardless of advances in AI," says Lemire. "The jobs will be different, even completely new, but there won't be a drop anytime soon because of artificial intelligence. Students who aspire to a career in the software industry must stay the course," he adds.
The efficiency of programmers, not their replacement, will be the great advantage of this revolution, according to Stephen Piccolo, professor at Brigham Young University (USA): "Surely some companies will hire less, at least initially, while understanding the implications of technology. But in the near future, this technology will make programmers more efficient rather than replacing them."
Another of the repeated aspirations that could allow AI is to program in natural language: ask ChatGPT in Spanish to write the code necessary to create an application, a website or a graphic. While something can be done, it will also be limited in scope for now. "Natural languages are full of ambiguities," says Emily Morgan, a professor of linguistics at the University of California Davis and co-author of a paper on how these systems encounter computer errors. "On the contrary, programming languages have to be unambiguous. That's why we want our programming languages to be unambiguous."
But you can order in Spanish something that ChatGPT converts into code. The problem then will be to find or iron out the problems: "We're going toward using natural language interfaces (like ChatGPT) to help generate a lot of our code," Morgan says. He explains, "You can ask ChatGPT for things like 'generate a template for a website,' but it will still generate code or HTML. It will need natural language to be translated into an unambiguous programming language."
You can follow EL PAĂŤS TecnologĂa on Facebook and Twitter or sign up here to receive our weekly newsletter.
Subscribe to continue reading
Read without limits
Read more
I'm already a subscriber