There won’t be an AI winter this time
Machine learning isn’t a “Skynet or bust” proposition
--
Disclaimer: My observations are influenced by the fact that I work on Cortex, an open source machine learning deployment platform.
Every few weeks, a new article predicting an imminent AI winter gets circulated. The arguments generally follow the same lines:
- The power of deep learning has been oversold to the public.
- We are farther away from artificial general intelligence than reported.
- Previous AI winters were caused by a similar hype bubble.
And while all of the above is true to an extent, the arguments miss something important: An AI winter similar to the one witnessed in the 80s, in which machine learning research and funding slowed to a crawl, simply isn’t possible anymore.
It’s true—we aren’t close to AGI or autonomous cars
For the last couple years, Starsky Robotics has been covered in major outlets as one of a few tech companies that would bring fully autonomous vehicles to the market. They had working demos, plenty of funding, and a talented team.
In March, Starsky Robotics shut down. In his post mortem (which you should read in its entirety), founder Stefan Seltz-Axmacher laid out the core reason very plainly: “supervised machine learning doesn’t live up to the hype.”
The hype he’s referring to is, of course, the endless promises from founders, journalists, and enthusiasts that technology like AGI and fully-autonomous vehicles are just months away. As Seltz-Axmacher reports, “Instead, the consensus (among researchers) has become that we’re at least 10 years away from self-driving cars.”
In the past, AI winters have been caused by similar cycles of exciting research leading to over-promises leading to disappointed investors and engineers giving up—but this time is different.
This time, even if the most grandiose promises of deep learning have not panned out, something else has happened:
Machine learning has become profitable.
Production machine learning is everywhere
Look at the most popular apps in the world:
- Netflix, YouTube, Facebook, Amazon, Instagram, Spotify, and TikTok all rely heavily on ML-powered recommendation engines.
- Snapchat, Instagram, and TikTok all use computer vision models to help users create, edit, and categorize visualize content.
- Gmail and Messenger both use NLP to enhance messaging for users—filtering spam, suggesting text, categorizing messages, etc.
- Google Maps, Uber, and Lyft rely on machine learning to calculate accurate ETA predictions.
That includes every single one of the most popular iOS apps.
These are the flagship products of the most valuable companies in the world—the same companies which, not coincidentally, are behind the majority of machine learning R&D. If you think any of these companies are going to stop investing in machine learning simply because they can’t build Skynet, you’re missing the point.
Production machine learning isn’t limited to tech giants, either. There are many startups who have already brought a product to market built on top of applied machine learning:
- Onfido uses machine learning to provide identity verification services to over 1,500 financial organizations worldwide.
- Ezra uses computer vision to provide full-body cancer screenings, currently operating in three states and growing.
- AI Dungeon operates a ML-powered text adventure game built on OpenAI’s GPT-2. They have over 1,000,000 players.
Within virtually every industry—medicine, agriculture, gaming, finance, security, etc.—there are companies who have successfully brought a machine learning product to market.
Machine learning isn’t a bet anymore
The reason that hype cycles were able to crash AI investment in previous decades was that AI, and by extension machine learning, were essentially bets.
Founders and researchers were speculating about a future in which machine learning might have commercial applications. When those bets didn’t pay off, the market collapsed.
Machine learning is no longer a speculative proposition, it is a widely-applied, commercially viable technology powering some of the most popular (and profitable) companies in the world. Google isn’t going to disband Google Brain or stop funding TensorFlow because Starsky Robotics and OpenAI—who represent some of the most ambitious technological projects in history—stumbled a bit.
Journalists who predict the Singularity will be here by Christmas might be wildly wrong, but they won’t be causing an AI winter.
