There are «Iron» — participate in the race.

22 октября, 2021 от Kinok Выкл

In the 2020, the balance of forces in the technological competition became extremely simply evaluated. The revolution of «deep learning of large models for large data» turned computing power into a key factor in the progress of almost all intellectually capacious industries: from the development of new drugs to new types of weapons. And where the AI is involved (and it is almost everywhere) computing power, in general, solves everything.

The superiority formula has become extremely simple:

The one who will have «bigger-more-faster» has the highest chances of winning in a technological race. And here everything resumes in the computing power of «iron» (HW) and algorithms (SW).

And with all due to all due to algorithms, but in this pair their role number 2. For the algorithm to invent, copy or even steal is still easier than HW. «Iron» is either there or it is not.

We were held in the USSR. It became even more critical factor in the era of «deep learning of large models for large data.»

Here are the two freshest examples.

For comparison, the GPT-2 language model, developed by OpenAI 2 years ago, struck the world by the fact that she had 1.5 billion parameters. And GPT-3, who had already published in 2020 had 175 billion parameters.

As you can see, models with trillion parameters are already given. And to learn them without years, and days, you need a «iron» of crazy computing power.

Those. You see, — there is «iron» — participate in the race, there is no «iron» — smoke aside.


Thanks for watching! Put husky I.