State of AI in 2021
Two tech investors, Ian Hogarth and Nathan Benaich, have been detailing the State of AI every year since 2018. They released their 2020 version in early October this year.
Hogarth and Benaich take a macroscopic look at the global artificial intelligence playing field, which they’ve summed up in a 177-slide report! (It’s a long but fascinating read for AI followers.) They define AI as:
“ A multidisciplinary field of science and engineering whose goal is to create intelligence machines.”
Before we jump into their 2020 assessment, let’s look at how they fared with their 2019 predictions.
In their 2019 State of AI report, Hogarth and Benaich made six predictions—4.5 of which were correct. The team accurately predicted:
The team was partially correct on one prediction: that Google will have a major quantum breakthrough and that five new startups will form with a focus on quantum ML. Of course, Google arguably achieved “quantum supremacy” in October 2019—that’s the correct half. Many quantum startups were launched in 2019, but fewer than five are focused specifically on quantum ML.
One prediction the team got wrong? AI governance. In the 2019 report, they predicted (hoped?) that governance for artificial intelligence would become a key issue, with at least one significant AI company making a big change to their governance model. Unfortunately, this did not happen.
This year’s report considers a few dimensions:
Here’s a summary of key findings in each area.
The buzziest area of AI today is natural language processing, or NLP. Next gen transformer language models, like GPT-3, will continue to unlock new NLP use cases, some of which we may even go on to use practically. But the concern here is the sustainability of such research. These systems are huge; GPT-3 uses hundreds of billions of parameters to run. Most companies can’t afford to live in this world of boundless compute power and endless associated costs.
This lack of access can significantly reduce competition—which isn’t good for research. Machine learning researches believe that progress in ML has stalled, leaving few organizations that can experiment significantly.
This concern plays out in the real world: many companies say they haven’t adopted big pre-trained AI Software like IBM Watson because they don’t have a true understanding of how it’s trained, how it will perform. This could easily introduce problems around ethics, reputation, and compliance—as we’ll see shortly.
Other research-related findings:
The U.S. maintains its reign of AI dominance, measured in acceptances of major academic papers. But if you look at the makeup of these researchers, more than two-thirds are born outside the U.S. Of AI researchers in America:
This indicates that AI talent is still drawn primarily to the U.S.—at least for now. Beyond the U.S., several new institutions around the world are dedicated to AI are formed (one of the team’s accurate predictions from last year).
Still, despite these findings, demand continues to outpace supply for AI talent. In 20202, academic brain drain—researchers leaving academic research—is acute, driven by increases in corporate recruiting. At least for now, this brain drain is having a negative impact on entrepreneurship.
Industry is the opposite of research—it’s what Hogarth and Beniath call the “task-specific” domain of AI. Despite more research breakthroughs seeming to come more often, it’s the on-the-ground AI that companies need. Of the fastest growing GitHub projects since July 2020, 25% focus on MLOps, short for machine learning operations. The investor team see this as a move away from R&D to operations—how to actually run models
Most businesses today that need an AI system actually need a not an overarching system, but one that can perform a very specific task and do so incredibly well.
For example, several companies have developed chatbots that outperform perform the research behemoth models like GPT-3 and BERT. These smaller models are important: they use much less compute power, making this type of AI adoption financially feasible for companies. NLP applications are being used widely, with implementation in Google Search and Microsoft Bing.
Other takeaways here:
Globally, we’ve reached the tipping point of technology. Internet users are becoming savvier and we can all point to a few instances where technology has hurt, not helped.
In artificial intelligence, we are beginning to realize ethical risks. Researchers have long warned about this black box problem of bias and accuracy, but it wasn’t until the Clearview AI scandal that facial recognition and data scraping became larger public concerns. Globally, some nations have begun passing laws that allow them to scrutinize when foreign parties takeover AI companies.
In response, some nations and U.S. states are starting to pass laws that limit certain technologies, including facial recognition. The next area ripe for regulation might be how algorithms work in other fields, like financial or insurance decisions.
Increasingly, companies are also becoming cogs of geopolitics. Semiconductor companies, crucial to the development of AI hardware, are unwittingly becoming geopolitical pawns. Some governments are even scrutinizing when foreign parties takeover AI companies.
The U.S. military, too, is experimenting with AI, likely adding capabilities to its strategies and techniques. The DoD—and thus the entire field of established defense contractors—is getting in on the AI action, working on projects ranging from intelligence analysis software to systems that could auto detect and disrupt electronic communications. The implications of this could be significant. (This ongoing experimentation might also explain why the DoD just released new regulations for data governance.)
With this background, we may increasingly see a new form of machine learning, one that can preserve and protect privacy. Known as federated learning, ML could offer a way for different parties to learn from the same data—without compromising privacy.
So, what does the team predict for the next 12 months? Of eight predictions, we think these are the most interesting:
With a strong track record, some of these predictions will certainly come true.
For related reading, explore these resources: