AIs were already good at finding the solution to a problem (the “what”). Now, this AI is designing the optimal process to get to that solution (the “how”). It is the difference between knowing that 2+2=4 and discovering the very concept of addition. Now, a study on Google DeepMind, published in Nature, represents a qualitative leap in the capabilities of AI. It is not just about optimizing, but about innovating in a primarily human field: the creation of abstract and efficient knowledge.
The fact that Google’s AI single-handedly rediscovered mathematical principles like Karatsuba’s (multiplying large numbers more efficiently) is monumental. He didn’t just remix existing code; he invented a better procedure, an achievement that had previously been the exclusive domain of exceptional human minds. Soviet scientists discovered this principle in 1960, after years of mathematical research. This AI rediscovered it from scratch, without having been taught.
Human algorithms are often based on intuitions and patterns we have discovered over decades. AI, freed from those biases, explored the space of possibilities more exhaustively and found methods that humans had overlooked. This suggests that our current algorithmic knowledge could be just the tip of the iceberg.
Imagine that a car engine, while driving on the highway, could redesign its own pistons to be more efficient. Or that a camera, while recording, would reinvent the way light hits the sensor to obtain a sharper image. It sounds like science fiction, but it is just the leap that AI has just taken: developing new algorithms that are more efficient than the best ones created by the human mind. In short, we are witnessing the moment when machines begin to rewrite the basic rules of their own operation.
Until now, AIs were excellent optimization tools. They could find the fastest route or recommend a song. But this new system, based on reinforcement learning (the same technique that AlphaGo used to defeat the world Go champion), has taken a much greater qualitative step.
Later, AI went further and designed novel algorithms for a fundamental operation in computing called “classification” (sorting data), which surpassed industry standards in efficiency. It’s like, After learning the rules of chess, an AI would not only become the best player, but would invent a new piece that would make the game more strategic.
And now, what, does it mean the end of programmers? The answer is a not resounding, but with a transformative nuance: it will not be the end, but it will be the end of programming as we know it.The job of the software engineer will not disappear, but will evolve from a line-by-line code creator to a problem architect and a supervisor of algorithmic geniuses. His new role will be more like that of an orchestra conductor than that of an individual musician.
The implications of this meta-advance are difficult to quantify. By improving the fundamental algorithms they use, from operating systems to databaseswe could experience a general acceleration of all technology.
Data compression processes could be more efficient, saving bandwidth and storage space. Complex scientific calculations could be performed in a fraction of the time. Phones and computers could become faster without changing their hardware, simply because the software they run would be better designed… by other machines.