What AI and power plants have in common

The story of artificial intelligence The (AI) development of the past five years has been dominated by scale. Huge progress has been made in natural language processing (NLP), image comprehension, speech recognition and more by taking strategies developed in the mid-2010s and leaving more computing power and more data behind. This has led to interesting power dynamics in the use and distribution of AI systems; one that makes the AI ​​look a lot like the grid.

For NLP, bigger is really better

The current state of the art in NLP is powered by neural networks with billions of parameters trained on terabytes of text. Keeping these networks simple in memory requires multiple advanced GPUs, and training these networks requires supercomputing clusters that are well beyond the reach of all but the largest organizations.

Using the same techniques, a considerably smaller neural network on significantly less text, but performance would be significantly worse. So much worse, in fact, that it becomes a difference in nature instead of just a difference in degree; there are tasks such as text classification, summary and entity extraction where large language models excel and small language models do not outperform chance.

As someone who has been working with neural networks for about ten years, I am genuinely surprised by this development. It is not technically clear that increasing the number of parameters in a neural network would lead to such a drastic improvement in capacity. However, here we are in 2022, training neural networks almost identical to architectures first published in 2017, but with orders of magnitude more computing power and better results.

This points to a new and interesting dynamic in the field. State-of-the-art models are too computationally expensive for almost any company – let alone an individual – to create or even implement. If a company wants to use such models, it must use one made and hosted by someone else – similar to the way electricity is created and distributed today.

Share AI as if it were a measured utility

Every office building needs electricity, but no office building can house the necessary infrastructure to generate its own power. Instead, they connect to a centralized power grid and pay for the power they use.

Similarly, a large number of companies can benefit from integrating NLP into their operations, although few have the resources to build their own AI models. This is exactly why companies have created great AI models and made them available through an easy-to-use API. By providing companies with a way to “connect” to the proverbial NLP power grid, the cost of training these large-scale state-of-the-art models is amortized across different customers, giving them access to this advanced technology. without state-of-the-art infrastructure.

To give a concrete example, let’s say a company that stores legal documents wants to display a summary of every document it has in its possession. They can hire a few law students to read and summarize each document individually, or they can use a neural network. Large-scale neural networks interacting with a law student’s workflow would dramatically increase efficiency in summarizing. However, training from scratch would cost a lot more than just hiring more law students, but if that company had access to a state-of-the-art neural network through a network-based API, they would plug you into the AI ​​”power grid” and pay for it. use of the summary.

This analogy has some interesting implications if we take it to the logical extreme. Electricity is a utility, as are water and transportation infrastructure. These services are so critical to the functioning of our society that in Ontario (from where I am writing) they are successfully maintained by crown companies (owned and regulated by the federal or provincial governments). These crown companies are not only responsible for infrastructure and distribution, but also for evaluation and quality assurance, such as testing water quality.

Regulating the use of AI is also key

In addition, like electricity, this technology can be misused. It has also been shown to have various limitations and potential abuse. Much has been learned about how these models could potentially cause harm through astroturfing and the spread of prejudice. Given how this technology will fundamentally change the way we work, governance and regulation are important to consider. Several providers of these NLP APIs have recently released a set of best practices for implementing these models, but this is obviously only a first step, building on this previous work.

Andrew Ng famous saying that “AI is the new electricity.” I think he meant that it will drive a wave of progress and innovation that will become crucial to the functioning of our economy, with the same impact as the introduction of electricity. The explanation might be a little hyperbolic, but it might be more appropriate than I initially thought. If AI is the new electricity, it will have to be turned on by a new set of power plants.

Nick Frost is co-founder of Coherent.