1) Language AI will take center stage, with more startups getting funded in NLP than in any other category of AI.
Language is humanity’s most important invention. More than any other attribute, it is the defining hallmark of our species’ intelligence.
Naturally, language pervades every facet of every business activity across every sector. The ability to accurately automate language therefore opens up virtually unbounded opportunities for value creation.
The field of natural language processing (NLP) has been upended and turbocharged in the past few years by a foundational new technology known as transformers, first introduced by Google researchers in a 2017 paper. We are only now reaching the point at which this dazzingly powerful technology is mature enough to be productized and commercialized at scale. A revolution in language AI, and thus in business, is around the corner.
Venture capitalists will plow record amounts of money into NLP startups in 2022. Leading NLP startups Hugging Face (last valued at $440M) and Cohere (last valued at $200M) will both become unicorns next year.
In the months and years ahead, expect a Cambrian explosion in NLP startup innovation as entrepreneurs identify a vast array of language-based activities across the economy to optimize, automate and transform using AI.
2) Databricks, DataRobot and Scale AI will all go public.
These three companies are among the first wave of big winners in the modern AI economy. They each provide tools and infrastructure to help other companies build AI, reflecting the common theme across technology cycles that infrastructure precedes applications.
All three companies boast astonishingly high revenue growth rates. All three raised big rounds in 2021 from “pre-IPO” investment firms, which typically invest in companies shortly before they go public: Databricks from Franklin Templeton; DataRobot from Altimeter and Tiger Global; Scale AI from Dragoneer, Greenoaks and Tiger Global.
Companies often make a high-profile CFO hire in preparation for an upcoming IPO. DataRobot announced this April that it had hired Damon Fletcher (formerly Tableau CFO) for the role. Databricks CFO Dave Conte, meanwhile, previously served as CFO of Splunk, where he took the company public in 2012. Don’t be surprised to see Scale AI make a high-profile CFO hire early in the new year.
3) At least three climate AI startups will become unicorns.
Climate tech has rapidly become one of the hottest categories in the world of startups, with record amounts of venture capital pouring into the sector this year. As previously explored in this column, opportunities abound for startups at the intersection of climate and artificial intelligence.
A number of climate AI startups have recently burst onto the scene with big funding rounds (despite limited commercial traction to date). Next year, a few of these players will ride the intensifying climate tech fervor to billion-dollar-plus valuations. The most likely unicorn candidates will be companies building tools for the new carbon economy (e.g., enterprise carbon accounting, carbon offsets infrastructure).
Potential unicorns: Cervest, ClimateAi, Gro Intelligence, Kettle, KoBold Metals, NCX, Pachama, Patch, Persefoni, Watershed
4) Powerful new AI tools will be built for video.
Video has become the dominant medium for our digital lives. Over 80% of all Internet data in 2022 will be video, according to Cisco. Every day, 7 billion videos are watched on YouTube and 100 million videos are uploaded to TikTok. From Netflix to Amazon Prime Video to Disney+ to Hulu to HBO Max and beyond, Internet streaming services’ user bases and content libraries continue to balloon.
And yet, compared to other data modalities like image and text, there has been relatively little focus to date on building deep learning-based products and capabilities specifically for video. This represents a massive market opportunity.
Expect to see a blossoming of AI tools for video in 2022, from video search to video editing to video generation. In the latter category, Synthesia’s $50M Series B raise earlier this month is a (both exciting and unnerving) sign of things to come.
5) An NLP model with over 10 trillion parameters will be built.
The field of natural language processing (NLP) today is defined by the development of ever-larger transformer-based models. This arms race will continue in 2022 (notwithstanding intriguing recent work from DeepMind on the power of smaller models).
In 2019, OpenAI’s GPT-2 became the first model with over 1 billion parameters (its 1.5 billion parameters seemed mind-bogglingly large at the time). In 2020, GPT-3 took the AI community by storm; with 175 billion parameters, it dwarfed everything that had come before it. But GPT-3’s reign as the largest AI model didn’t last long. In 2021, the trillion-parameter barrier was broken by models from Google (1.6 trillion parameters) and the Beijing Academy of Artificial Intelligence (1.75 trillion parameters).
Expect this hockey-stick growth in the size of large language models to continue next year. There is a good chance that 2022’s largest model will come from OpenAI and be named GPT-4.
6) Collaboration and investment will all but cease between American and Chinese actors in the field of AI.
It is no secret that geopolitical tensions between the United States and China are ratcheting up, with cutting-edge technologies like artificial intelligence representing a particularly contentious touchpoint in the conflict. This will get worse—much worse—in 2022.
In just the past few weeks, the U.S. government added AI startup SenseTime, drone company DJI, and several other leading Chinese AI organizations to an investment blacklist. These are among the most important AI companies in China.
The Committee on Foreign Investment in the United States (CFIUS) is acting increasingly aggressively to prevent Chinese organizations from investing in or gaining access to U.S.-based AI technology. The influential National Security Commission on Artificial Intelligence (NSCAI), chaired by Eric Schmidt, has further fanned the flames of an AI arms race with China, for instance encouraging the U.S. government to wall off U.S. university research in AI from the Chinese.
The upshot of all this: as 2022 progresses, it will become all but impossible for American and Chinese actors—entrepreneurs, investors, corporations, business leaders, academic researchers—to meaningfully work together on AI initiatives.
7) Multiple large cloud/data platforms will announce new synthetic data initiatives.
Getting the right data is the most important and the most challenging part of building AI products today. Synthetic data offers compelling advantages over the status-quo approach of collecting and labeling real-world datasets.
Gartner has predicted that by 2024, synthetic data will account for 60% of all data used in AI development. Facebook’s acquisition of synthetic data startup AI.Reverie two months ago is a canary in the coalmine.
Next year, multiple major computing platforms will launch new synthetic data efforts as they recognize the importance of this technology to tomorrow’s AI stack and seek to attract more builders to their ecosystems.
Likely candidates: Amazon Web Services, Microsoft Azure, Google Cloud Platform, Unity Technologies, Scale AI
8) Toronto will establish itself as the most important AI hub in the world outside of Silicon Valley and China.
It is not an exaggeration to say that modern artificial intelligence was invented in Toronto, thanks to the work of deep learning pioneers like Geoff Hinton. Though it generates less buzz than other geographies, Toronto remains one of the most important AI hubs in the world.
It is swarming with AI talent. According to a recent CBRE report, the Toronto-Waterloo metropolitan area is the #2 largest market for technology talent in all of North America (behind only the Bay Area)—and the #1 fastest-growing. The Vector Institute, co-founded by Hinton in Toronto, is one of the largest AI research organizations in the world. From Google to Microsoft to IBM, the world’s largest tech companies have established major presences in the city in recent years.
Historically, Toronto has had a reputation as a top-notch AI research hub but a comparatively underdeveloped startup ecosystem. This is changing fast. Ada (chatbot platform), Cohere (NLP), Deep Genomics (AI for drug discovery) and Waabi (autonomous vehicles) are just a few examples of Toronto-based AI startups that have raised monster funding rounds in recent months.
Expect more world-class AI startups to emerge from Toronto in the coming year.
9) “Responsible AI” will begin to shift from a vague catch-all term to an operationalized set of enterprise practices.
AI technology is improving faster than is our ability to deploy it responsibly, ethically and equitably.
A growing movement has emerged to advocate for the responsible use of AI, led by researchers like Timnit Gebru, Joy Buolamwini and Cathy O’Neill. This push for more responsible AI spans a broad set of issues including AI bias, data provenance, model explanability and model auditability.
While awareness of these issues is growing, the topic remains sufficiently abstract that, by and large, AI practitioners do not build “responsible AI” practices into their day-to-day workflows.
2022 is the year that this will begin to change, as responsible AI practices and toolkits become productized and operationalized. These products will come both from tech giants (e.g., Microsoft, IBM) and from newer startups (e.g., Parity, Fiddler Labs). Over time, responsible AI practices will shift from “nice-to-have” efforts within forward-thinking organizations to standard practice across industries.
Regulation will provide an important impetus: see, for instance, the E.U.’s proposed Artificial Intelligence Act and New York City’s new law mandating audits for companies that use AI in hiring decisions (the first of its kind). Corporate efforts to self-regulate will also move the ball forward here. Just this month, a group of Fortune 500 companies including Walmart, Nike, General Motors and CVS announced the Data & Trust Alliance, a cross-industry consortium whose stated goal is to “detect and combat algorithmic bias.”
10) Reinforcement learning will become an increasingly important and influential AI paradigm.
The dominant approach to AI today is supervised learning, which entails collecting a lot of data, labeling it, and feeding it into an AI model so that the AI learns useful patterns about the world. Unsupervised learning, a similar approach but without the need for human-generated labels, has also begun to gain traction in recent years.
But there is another paradigm in AI, which has been around for decades but whose vast real-world potential is only starting to become clear: reinforcement learning.
In reinforcement learning, the AI is not trained on historical real-world data; it is not given the “answer key” and told what to pay attention to, as in supervised learning. Instead, it is allowed to open-endedly explore its environment, learning about the world as it goes, guided only by a particular objective that it seeks to optimize for.
Reinforcement learning powered DeepMind’s landmark AlphaGo triumph. Increasingly, it is being used by researchers and startups at the bleeding edge of AI to unlock unprecedented AI capabilities, from recommendation engines to robotics to autonomous vehicles and beyond.
Reinforcement learning may offer a path to a more sophisticated, flexible form of machine intelligence. In a provocative paper published a few months ago, DeepMind went as far as to posit that reinforcement learning, by itself, could take us all the way to “artificial general intelligence.” As the most advanced AI research organization in the world, DeepMind is worth paying attention to.
See here for our 2021 predictions (published December 2020). See here for our retrospective analysis of those 2021 predictions (published December 2021).
Note: The author is a Partner at Radical Ventures, which is an investor in ClimateAi, Cohere and Waabi.