Every year, Nathan Benaich and Ian Hogarth publish a report on “The State of AI” that examines cutting-edge research, how A.I. is being applied now, and the politics and regulation of the technology.
Benaich, who is the principal in the early-stage investment fund Air Street Capital, and Hogarth, a prominent London angel investor, bring a broad perspective to the yearly status check, which is out today. I asked them what they saw as the year’s most important developments. Here’s some of what they highlighted:
Organizations are starting to trust A.I. to run critical operations, not just trim costs or improve sales at the margin. As an example, Benaich points to Ocado, the British online grocery that also sells its technology to other grocers globally, including Kroger in the U.S. Its A.I. software is so good at forecasting demand for 55,000 items that it can be trusted to automatically make decisions on stock replenishment in 98% of cases.
China is on track to overtake the U.S. in A.I. research. Hogarth says the amount of A.I. research from China, and its improving quality, are striking. “They’ve gone from publishing no research papers on A.I. in 1980 to the largest volume of A.I. research now,” he says. Increasingly these papers are appearing in top-ranked academic journals. “They are blitzing past Western universities that have been at it a lot longer,” he says.
A.I. is poised to revolutionize the biological sciences. The report notes advances such as DeepMind’s AlphaFold 2, which can accurately predict many protein shapes based on their genetic sequences. Benaich also highlights the recent development of ARES, deep learning software, created by a team at Stanford University and a startup called Atomic AI, that can predict the three-dimensional structure of strands of messenger RNA. Both developments may transform large swathes of biological research, including the quest for new drugs.
Transformers are becoming ubiquitous. A transformer is a kind of neural network that is good at spotting statistical patterns in long sequences. They’re the driving force behind the advent of ultra-large language algorithms like OpenAI’s GPT-3 that exhibit uncanny performance at a range of language tasks, such as writing, translation and answering questions. But they are also the key to A.I. that can master many computer-vision tasks, such as classifying objects.
Everyone is building their own ultra-large language A.I. The stunning success of GPT-3 has convinced a growing number of people to develop their own ultra-large language models. (Many companies and governments are wary of being dependent on GPT-3’s creator, OpenAI, a profit-oriented company closely affiliated with Microsoft.) Benaich and Hogarth spotlight some of these efforts. One is A.I. research collective EleuthrAI that is building an open-source version of GPT-3.
Every year, Benaich and Hogarth also make predictions for the year ahead. Among last year’s forecasts was that an A.I.-based drug discovery would either have an initial public offering or be acquired for at least $1 billion. In this prediction, the co-authors were right twice over: Exscientia, a British company that has three medicines discovered by its A.I. techniques currently in early-stage human clinical trials, was valued at more than $3 billion in a NASDAQ IPO. And Utah-based Recursion Pharmaceuticals, which uses A.I. combined with high throughput robotic laboratory equipment to search for new drugs and has four in clinical trials, also went public at a $3.1 billion valuation.
Another prediction was that Nvidia’s takeover of British chipmaker ARM would be blocked by regulators, either in Europe or the U.S. Well, so far, Benaich and Hogarth are half-right on this one. The buyout has not yet gone through as it has encountered regulatory probes on both sides of the Atlantic. But neither has the deal been definitively quashed.
So what are Benaich and Hogarth predicting for the next 12 months? Here are a few:
•At least one of the five most prominent startups focused on producing computer chips specialized for A.I. applications—Graphcore, Cerebras, SambaNova, Groq, or Mythic—will be acquired by a larger semiconductor maker or by a big tech company.
•The market value of ASML, the Dutch company that makes semiconductor manufacturing equipment, will reach $500 billion.
•The new A.I. research organization Anthropic will publish the results of breakthrough experiments that will position it as “the third pole” in efforts to develop artificial general intelligence (AGI), alongside existing players DeepMind and OpenAI.
***
Before we get to this week’s news, a reminder that if you want to know more about how A.I. can transform your company and your industry, apply to attend Fortune’s inaugural Brainstorm A.I. conference in Boston, November 8-9. The conference will offer a deep dive into the business applications of A.I. Among those scheduled to speak are Moderna Co-founder and Chairman Dr. Noubar Afeyan, who will discuss the promise of A.I. in healthcare; Stanley Black & Decker CEO Jim Loree will explain how machine learning is turbocharging his business. Amazon senior vice president and head scientist of Alexa Artificial Intelligence Rohit Prasad will explain how A.I. can help companies tailor their services to individual customers. Meanwhile, Levi Strauss senior vice president, and chief strategy and artificial intelligence officer Katia Walsh will share her secrets about using machine learning to better engage with customers. I hope to see you there!
***
Jeremy Kahn
@jeremyakahn
jeremy.kahn@fortune.com
A.I. IN THE NEWS
The U.S. has already lost the A.I. technology race to China, says former Pentagon official. Nicolas Challian, who had been the Pentagon's chief software officer until resigning earlier this month, told The Financial Times that he left out of frustration with the U.S. military's slow pace of technological transformation. He particularly faulted the U.S. for not responding to Chinese cyber threats and for failing to keep pace with China's rapid integration of A.I. into its military. “We have no competing fighting chance against China in 15 to 20 years. Right now, it’s already a done deal; it is already over in my opinion,” he told the newspaper.
Controversial facial recognition company Clearview AI says it's using a massive trove of data to enhance its capabilities. Hoan Tan-That, the company's CEO, told Wired that the company now has a dataset of more than 10 billion faces harvested mostly from public social media accounts. He said it is using this massive data trove to further enhance the ability of its facial recognition algorithms, training them to sharpen blurry images and even to recognize people wearing masks.
U.S. officials propose updated bill of rights in response to the growing use of A.I. The top two officials at the White House's Office of Science and Technology Policy, Eric Lander and Alondra Nelson, co-wrote an op-ed in Wired calling for legislation that would guarantee Americans certain protections from the potential harmful effects of A.I. Among the rights they propose are the right to know when an A.I. system is making a decision that could impact your civil liberties, and freedom from being subjected to A.I. that hasn't been audited to ensure it is accurate and unbiased.
DeepMind turns a profit. The London-based A.I. research company, which is owned by Google-parent Alphabet, said in financial filings that it had posted its first-ever profit in 2020. The company said it made $59.8 million on revenues of about $1.13 billion compared to a loss of $650 million on revenues of $362 million the year before. But all of DeepMind's revenue comes from internal royalty payments from Google and other Alphabet companies for use of its technology and it isn't entirely clear how the value of those payments is established or why they jumped so significantly from the year before. CNBC has more on the news here.
EYE ON A.I. TALENT
Arize AI, a Berkeley, Calif., company that helps businesses understand and monitor their A.I. systems, has hired Amber Roberts to be its sales engineer for machine learning, according to a company blog post. Previously, she was a product manager for A.I. and machine learning at data analytics software provider Splunk.
DataKind, a New York-based nonprofit that helps connect data scientists with projects that have social impact, often for other charities and non-profits, said it has hired Lauren Woodman as its new CEO. Woodman, a former Microsoft executive, was previously CEO at NetHope, a Seattle-based organization that helps nonprofits adopt innovative technology.
Accrete, a New York-based company that makes A.I. software for intelligence analysis, announced it has hired Brian Drake to be chief technology officer for its federal contracting subsidiary, Accrete AI Government. Drake had been chief of the senior leadership and technology team at the U.S. Defense Department, and before that served in a variety of roles at the Defense Intelligence Agency.
EYE ON A.I. RESEARCH
Another big step in protein structure prediction. Proteins are the building blocks of life. And much of their function is determined by their form—the three-dimensional shape that a protein takes. Knowing a protein's structure can be crucial in drug development. So DeepMind's breakthrough late last year in predicting the shape of proteins just from the genetic sequences that encode them, was a big deal. But, some scientists grumbled, a lot of the really important questions in biology are about the interactions between proteins, not just the shape of a single protein. And DeepMind's AlphaFold algorithm wasn't designed to predict these.
First, a number of researchers tried to essentially hack AlphaFold into predicting protein complexes and managed to get decent results. But now the DeepMind team has taken AlphaFold and specifically retrained it on a database of protein structures and again managed to outperform all other approaches. The company said its new system could predict 67% of the contact points between two dissimilar proteins, with what it called "high accuracy" achieved in 23% of cases. When the two proteins were largely similar, it could predict the contact points in 69% of cases, and achieve high accuracy 34% of the time. While this kind of accuracy is not yet at the level AlphaFold has for single proteins, it is still likely to be useful for biologists and, given how quickly DeepMind made this new advance, it's also likely further improvements will come in the next 12 months.
FORTUNE ON A.I.
VMware CEO on cloud push, 5G, and artificial intelligence—by Jonathan Vanian
Tesla to move headquarters to Texas, in latest loss for California’s tech industry—by Rey Mashayekhi
Vehicle-to-grid technology could ease Europe’s energy transition woes (Commentary)—by Katrin Zimmermann
Google to start nudging people to save the planet—by Robert Hackett and Declan Harty
BRAIN FOOD
Could A.I. help prevent air disasters? The fiery crash of an Air France Concorde flight immediately after takeoff from Paris in 2000, killed 113 people and helped bring about the end of supersonic commercial aviation. (Of course, some companies are now trying to revive it.) Investigators later determined that a piece of metal had fallen off a plane that had taken off just before the Concorde. This metal scrap blew out the Concorde's tires and the exploding debris ruptured the Concorde's fuel tanks, leading to the catastrophe. While it was a freak accident, it was hardly unprecedented. Some aviation industry sources say that this kind of runway debris causes $4 billion to $13 billion in damage to aircraft each year. Could A.I. prevent such an accident from ever happening again?
University of Nebraska researchers hope so. They have created a dataset of 31 foreign objects, from batteries to fuel caps, that may find their way on to runways. The dataset includes more than 30,000 images of these objects taken on runways in different weather and light conditions. The hope is to create a machine-learning algorithm that could automatically detect dangerous objects before a plane rolls over them at high-speed. In the future, a drone, or maybe even a stationary camera positioned at the right angle, could survey a runway and determine whether it is clear of any hazards. It is the kind of vital, but boring, task that humans often struggle to perform consistently. That makes it a good candidate for A.I.
