So, AI? Where Is It Going?
by Bruno Accioly – 11.10.2025
Introduction
If you follow the news about Artificial Intelligence, you may be confused. One day, headlines say that AI is a fraud, that it only repeats nonsense, that it will never replace human reasoning. The next, they warn that it will eliminate jobs, revolutionize industries, and transform our civilization forever. So, are we facing an illusion or a revolution?The truth is that Artificial Intelligence did not emerge yesterday. Since the 1950s, researchers have tried to create systems capable of learning, reasoning, and solving problems. There have been cycles of enthusiasm and disappointment — the so-called "AI winters" — spanning generations. But the decisive leap came in 2017, when a Google team introduced the Transformer architecture, the foundation of current models such as GPT, Claude, and Gemini. In other words, we are talking about a technology that, in fact, has only eight years of real maturation.And in that short interval, the advances have been astonishing. One reason is that these systems do not merely perform tasks: they help the engineers themselves improve them, acting as accelerators of scientific and technical research.But there is also a human factor: everyone has an opinion about intelligence. Just as in Design, where anyone feels comfortable criticizing or offering an opinion even without training, Artificial Intelligence awakens the same impulse. After all, we are all endowed with intelligence, so we feel we "own" the concept. This generates heated debates — from philosophy, psychology, computer science, and economics — as well as, of course, the spontaneous opinions of the public at large.The point is that, between hype and skepticism, the reality of AI is more interesting than either extreme. We are not facing an empty trick, nor a divine oracle. We are facing a young, powerful technology that is still taking shape — one that may say as much about us humans as it does about machines.
AI as a productive tool in the right hands
If in the first part we saw how naive or pseudoskeptical doubt tends to underestimate Artificial Intelligence, it is now important to look at the other side: what is actually already happening when it is used competently.
The metaphor is simple: a drill, in the hands of a child, can be a threat. But in the hands of an experienced carpenter, it is a powerful instrument of creation. The same applies to latest-generation language models (LLMs). Used poorly, they can generate incorrect or misleading answers. But when employed by people who know what they are looking for, they already become a major asset for accelerating productivity, reducing costs, and expanding creative horizons.Today, we have solid evidence of this:
- In science, researchers such as Scott Aronson have published papers in which GPT-5 suggested critical steps in complex mathematical proofs — not replacing the scientist, but accelerating the discovery process.
- In programming, engineers who use tools such as GitHub Copilot or corporate AI platforms report gains of up to 30% in development speed and code review, making it possible to release faster and safer versions.
- In customer support and services, controlled studies show that call center agents achieved an average 14% increase in productivity with AI support, reaching 35% among the least experienced. AI does not eliminate the human; it raises the overall level of quality.
In general intellectual work, experiments conducted by universities indicate increases of up to 37% in the execution of writing, planning, and synthesis tasks when professionals have the support of LLMs.
But perhaps what matters most is not in the percentages, but in the logic that emerges: the more execution is automated, the more the value of conception and human judgment grows.
AI is fast at implementing, but only people know how to discern what is worth implementing. It accelerates the production of alternatives, but it is up to us to choose which paths make sense, which opportunities are valuable, and which risks should be avoided.
Breakneck Evolution in 8 Years
| Year | Achievement / Event | Context / Details | Importance |
| 2017 | AlphaGo defeats Ke Jie (3–0) | AlphaGo Master defeats the world Go champion. | A symbol of AI's mastery of strategic games previously considered unreachable. |
| 2017 | AlphaGo Zero / autonomous mastery | A version without human data learns on its own and surpasses all previous versions. | The first clear demonstration of large-scale superhuman self-learning. |
| 2021–2022 | Hutter Prize (text compression) | Progressive improvements in the compression of large Wikipedia corpora. | An indicator of advances in language efficiency and representation. |
| 2023 | Programming benchmarks (SWE-bench) | Early models solve ~4.4% of real-world software tasks. | Shows the difficulty of extrapolating to practical coding problems. |
| 2024 | Improved SWE-bench (71.7%) | New models achieve a massive leap in the benchmark in just 1 year. | Demonstrates unprecedented acceleration in autonomous programming capabilities. |
| 2024 | AlphaProof reaches silver-medal level at the IMO | AI system solves mathematical olympiad problems, scoring like a silver medalist. | The first serious entry into elite human mathematics competitions. |
| 2025 | Gold medal at the IMO (Google/OpenAI AI) | Models solve 5 of the 6 problems at the International Mathematical Olympiad. | A symbolic milestone: AI reaches the top level of elite school mathematics. |
| 2025 | AI at IPhO gold-medalist level (Physics Supernova) | A score of 23.5/30 points, top 14 worldwide. | Proof that systems already rival humans in theoretical physics. |
| 2025 | AIxCC (DARPA Cyber Challenge) | AI tools find vulnerabilities in millions of lines of code, with US$4 million in prizes. | Demonstrates practical application in critical cybersecurity. |
| 2025 | SafeBench | AI safety benchmark/competition with significant prizes (US$50k+). | Establishes new standards for model evaluation and robustness. |
| 2025 | GPT-5 assists with quantum complexity proof (Scott Aronson) | An arXiv paper on the limits of techniques in QMA acknowledges that a crucial technical step came from iterations with GPT-5. | The first time an AI contributes directly to cutting-edge theoretical mathematics research. |
Trajectory of Achievements
The trajectory of artificial intelligence's achievements since 2017 shows a clear shift from the symbolic to the substantive. After the impact of AlphaGo defeating world champions at Go and demonstrating the power of autonomous learning, AI began to prove itself in more abstract fields: compression, language benchmarks, and, more recently, programming. Between 2023 and 2024, the leaps in benchmarks such as SWE-bench marked AI's entry into the territory of practical software engineering. During the same period, AlphaProof showed that mathematical olympiad problems were no longer out of reach.In 2025, the narrative took on historic proportions: a gold medal at the IMO, elite performance at the International Physics Olympiad, and victories in cybersecurity challenges such as AIxCC and SafeBench. Most remarkable, however, was the episode in September 2025: a Scott Aronson paper in complexity theory acknowledged that a central technical step in the proof had been suggested by GPT-5. Unlike winning games or solving formatted exercises, here AI acted as a scientific research partner, accelerating the discovery of a new result.The arc is clear: from victories in board games to prizes in science, mathematics, and security, and ultimately to active collaboration in frontier research. The year 2025 is cemented as the moment when AI models ceased to be merely systems that outperform human benchmarks and began to become coauthors of scientific and intellectual production itself.
The false comfort of denial
We need to stop trying to discredit Artificial Intelligence.
Not out of "respect" for the companies developing it; they do not need volunteer advocates. Nor because AI might be "offended" by criticism. But because this neo-Luddite attitude creates a tension that clashes with reality. The result is a discourse that mixes irrational fear with a veneer of wisdom, like someone who wants to appear lucid but actually repeats falsehoods to protect themselves from what they do not understand.
The historical contrast is glaring.In just eight years, since the publication of the Transformer architecture in 2017, neural networks have gone from mediocre translators to models capable of:
- comfortably surpassing the Turing Test, to the point of making it obsolete as a metric;
- achieving top scores in complex benchmarks, such as mathematics and physics competitions;
- winning prizes and taking first place in fields once considered "inaccessible" to machines.
In 2010, I remember IT colleagues who mocked AI. They said: "It will never pass the Turing Test." Well, it not only passed, but surpassed it. And suddenly the deniers changed their tune: "Oh, but that doesn't matter anymore." It is curious how the bar of disbelief keeps moving to ensure that AI remains "beneath humans and never this or that," regardless of the growth it has already demonstrated.Today, while uninformed skeptics chant their litany — "it's just snake oil, a giant autocomplete, it will never have genuine intelligence, it's a stochastic parrot, it has reached a plateau, it will hit a wall, it will steal jobs" — reality advances in silence:
