Evolution instead of revolution: what slows down the development of artificial intelligence

Singularity is a term that can often be heard in discussions of the future predicted for artificial intelligence. At this point in its development, “machine intelligence” will surpass human intelligence and will be able to create its own kind.

For many scientists, this prospect is of grave concern. After all, it is unknown what the AI ​​will decide to do with humanity. At the same time, according to many experts, this prospect is a matter of the very distant future.



However, such optimism is not inherent in everyone. In particular, Leopold Aschenbrenner, a former Open AI employee who was fired in April of this year for disclosing confidential information, published a very remarkable report. In it, he warns humanity about the imminent creation of AJI - a universal AI that will be able to immediately perform all tasks available to certain modern types of “machine intelligence” (texts, images, videos, etc.).

The danger is Technology, according to Aschenbrenner, is that she will think for herself and use all the skills available to her to accomplish a goal that only she knows.

But that's not all. Following AJI, ASI will appear - that same singularity. At the same time, the creation of universal AI, according to Aschenbrenner’s calculations, could happen by 2028. ASI, in turn, will be created by AJI itself within a few months.

A scary prospect...

However, everything is not so fatal if we take into account the factors that will stand in the way of such a rapid development of AI.

Firstly, neural networks are quite “gluttonous” and every year for their evolution they require more and more computing power. Today they practically reach the “ceiling” accessible to modern computers.

Secondly, neural networks lack data. The AI ​​has already “browsed” the entire Internet, which means that its further learning will occur much less rapidly.

Thirdly, AI consumes quite a lot of electricity. According to experts, by 2028, AI will absorb an amount of energy that would be enough to supply an entire American state.

Finally, fourthly. Экономическая validity. As of 2023, the revenue of the entire AI industry was $3 billion. At the same time, the total cost of creating and training neural networks has already exceeded 50 billion.

Thus, despite Aschenbrenner’s alarmist statements, artificial intelligence is more likely to undergo gradual evolution rather than revolution.