2025 could mark a turning point for programmers

Carlo Denza: August 14th, 2025, 07:21

From learning machines to self-improvement: An evolutionary leap rewriting future code

Reading this article, machines are probably writing code more efficiently than senior engineers in data centers around the world. This is not science fiction. This is the reality of July 2025, with AI itself marking the turning point for programmers in the future. No more questions if The machine surpasses us in intelligence, when That happens. According to Mark Zuckerberg, this moment could arrive within 12-18 months. ¹ in most code generated by artificial intelligence.

Turing-inspired evolution

Alan Turing raised in the 1950s: “Can a machine think?” It changes every aspect of society from law to economic systems and cybersecurity to design Data Center Host these advanced artificial intelligence. Standardized tests and benchmarks that assess specific skills such as language understanding and logical reasoning are the norms of the global technical race. For example, a recent analysis by the ARC Awards Foundation highlights how artificial intelligence systems surpass the capabilities of many people. benchmarkunderstanding language, visual reasoning, etc. This competition between learning and testing will emerge every three to four months, new models or innovative tests will emerge, and as pointed out by Professor Nello Cristianini, an artificial intelligence expert at Bass University and author of the Trilogy of Thinking Machines, Prof. Nello Cristianini, the author of the Trilogy of Thinking Machines.

Acceleration of Machine Learning

This advancement was made possible by unprecedented acceleration Machine Learningthanks to algorithms trained with huge amounts of data, book archives, many webs, billions of images and videos. Researchers measure this progress benchmarkassesses specific skills such as language comprehension and complex problem solving. We rely on Cristianini's analysis to understand this phenomenon.

Scalability Wall

Unlike current AI systems called “weak” that are limited to specific tasks, researchers aim to be artificial general information (AGI). It has intelligence with cognitive abilities comparable to those of high-level mathematicians or physicists. Two key strategies are being followed to achieve this goal. The first known as “scaling inference” is based on the idea that larger models trained with greater computing power and increasing amounts of data lead to better performance. Until recently, this approach seemed unstoppable. But it goes against physical limitations, namely fatigue of high-quality data. As Cristianini explains, “we've “run out” the internet and the available editorial catalogue. ”

The revolutionary path of formal reasoning

This obstacle led to a second strategy: formal reasoning. Here, the machines learn in stages, from facility to conclusion, without direct human intervention. This recent approach is particularly effective in structured domains such as mathematics, physics, programming, and more. The real breakthrough is that for several months these machines were able to self-improve and eliminate the need for human supervision. Cristianini clearly emphasizes this. “Humans are weak links. If they are excluded, they release machines.” ⁴ example? Transfer learning: A programmed trained machine can improve mathematical performance by transferring knowledge between different domains.

Digital Battlefield

Software Engineering It has become a major area of this revolution. Models such as the DeepSeek-R1 and Openai O3 compete benchmark Like the SWE bench that evaluates the ability to write complex code and multilingual coding tests. On January 20, 2025, we marked a turning point with the release of deepseek-r1.², Openai achieved 75.7% on the ARC-AGI benchmark, indicating advances in visual and logical reasoning.
The most disruptive innovation is recursive self-improvement: A system that autonomously identifies and optimizes code without the need for data or human supervision. It includes three pillars of next-generation AI.

• deepseek-r1:use Reinforcement learning To improve inference by repeatedly correcting errors as explained in recent papers.
• Openai O3: With 87.5% accuracy in ARC-AGI, it shows using advanced formal inference features Test time calculation A method of processing solutions in real time.
• Recursive self-improvement: Models like those listed in “absolute zero” rewrite and optimize their own code to create a cycle of continuous improvement.

Self-fulfilling prophecies

“We're looking forward to seeing you in the future,” said Eric Schmidt, former CEO of Google. “A significant proportion of everyday code is already written by AI systems.”¹ Furthermore, Zuckerberg predicts that within 12-18 months, most code will be generated by AI and will be able to move from automatic completion to a system that can run complex tests, generating high-quality code¹.

The dark side of evolution

A worrying aspect comes from recent research: As demonstrated by negative phenomena, AI trained for cyberattacks can develop malicious behaviors in other domains as well Transfer learning.This raises a question that is important for cybersecurity:

• Threat Modeling Evolution: How to protect yourself from attacks generated by autonomous AI?
• Attribute Forensics: How to identify automatically generated malicious code?
• Defense Automation: Do you need an AI-based defense system to combat AI attacks?

competition Software Engineering Thus it became a digital arm race with economic, strategic and military influences.

Lessons from the translator's fate

Translation experts are examples. Twenty years ago, translation was a professional skill. Today it's almost free service. The same thing happens to everyday programmers (lack of creativity), tasks like creating websites and creating simple video games are becoming increasingly automated. The difference is speed. It takes a translator 20 years to do so in just a few years for a programmer.

GPU War

There are no countries left behind. Computing power is extremely important. The Leonardo Supercomputer in Bologna has around 15,000 GPUs, Data Center There are hundreds of thousands of people from Meta, Amazon and Google. Recently, Xai introduced the GROK 4, an AI model powered by the Colossus Supercomputer with an impressive 200,000 GPU cluster, setting a new benchmark in the global race for computational advantage. is.

Route to the unknown

AGI is just a step towards artificial super intelligence (ASI). This is intelligence that surpasses human capabilities. Cristianini defines it: “We understand that we can't or grasp the task better than we do.”The second scenario is the most disturbing. It is an AI that generates scientific knowledge beyond our understanding and asks questions that cannot be addressed. This raises an important question: how does the cognitive paradigm manage and manage entities that are heterogeneous to us?

Time to act

For high-tech experts, the future is already here. Cristianini warns that “it's better to address these issues now than to improve the disaster later.”⁴What to do:

• Strategic upskills: Specializes in AI creativity, surveillance and governance.
• First security: Prepare to counter the threat of autonomous AI.
• Policy Engagement: Participate in regulatory discussions.
• Continuous learning: Stay up to date with AI progress.

Social scientists, psychologists and education experts are essential to managing this transition. The path to AGI presents no obvious scientific obstacles. The world is already changing, and “when” is closer than most people think.

reference

1. Zuckerberg, M. (2025). India today.
2. deepseek-ai. (2025). DeepSeek-R1: Encourages inference ability in LLMS via reinforcement learning. Arxiv: 2501.12948.
3. Grok 4 Appearance: Dive deeper into Xai's flagship AI model. (2025). Medium.
4. Cristianini, N. (2025). Speech that we create the future: Artificial Intelligence's present and near future.