In 2012, Tom Davenport, well-known thinker on management and technology, and D.J. Patil, the computer scientist who would later serve as chief data scientist in the Obama Administration’s Office of Science and Technology Policy, penned a provocative Harvard Business Review essay in which they argued, as the title of the article stated, that data scientist was the “sexiest job of the 21st Century.”
But Joe DosSantos, chief data and analytics officer at data analytics software company Qlik, hates their conclusion. “The essay created unreasonable expectations of what a data scientist can do,” DosSantos says, adding these over-hyped expectations worsened the war for talent by encouraging companies to look for “unicorn” candidates who possess all of the skills Davenport and Patel described.
Instead of hunting for data scientists who are the masters-of-everything that Davenport and Patil value, DosSantos says, companies should be thinking about how to build teams in which data literate analysts work alongside subject-matter experts from various business units. “Data science is a team sport,” he says. “And data scientists are arguably not even the most important people on the team.”
Developing effective algorithms for companies requires many steps. The most important of which, DosSantos says, are “aligning the organization around an important problem to solve and then figuring out how an algorithm might help solve that problem.” At both of those things, he argues, most data scientists are “useless.”
Other important steps include thinking about what data is available and thinking through the ethics of marrying that data with a particular business case. Then a company has to obtain the data, clean it, build the algorithm, and train it on the data. And then it must test the algorithm, and figure out how to deploy it. Finally, it must monitor the algorithm’s performance to ensure it is continuing to work as designed. Only a small number of those tasks absolutely requires a data scientist, DosSantos says: building the algorithm and training it. Subject matter experts or people with different kinds of engineering and analytic backgrounds can handle most of the others.
DosSantos says that it’s also important to distinguish between tasks that require a data scientist, like designing an algorithm and designing ways to monitor the algorithm once it is in production, and those that merely need a good data engineer, such as the creation of clean data sets that will be used to train the algorithm. “There is a kind of Batman and Robin teamwork between the data scientist and the data engineer,” he says. At Qlik, he says, for every data scientist working on a project, there are two data engineers.
And with that, here’s the rest of this week’s news in A.I.
A.I. IN THE NEWS
A.I. chipmaker Graphcore announces a new 3D chip and a roadmap for the world's most powerful A.I. supercomputer. The company, based in Bristol, England, unveiled a new A.I. chip, called Bow, that is built vertically, stacking two silicon wafers on top of one another. The new chip is able to run 40% faster while using 16% less energy than its predecessors, according to a story in tech trade publication IEEE Spectrum. The company also unveiled a roadmap to create an ultra-massive 10 exaflop supercomputer capable of running A.I. systems that process 500 trillion variables at once by 2024. The so-called "Good Computer," which will cost $120 million, is twice as powerful as a rival system that Facebook plans to build.
Intel's self-driving car division, Mobileye, files confidential paperwork for its IPO. That's according to Reuters, which says the listing is poised to be one the year's biggest public offerings. Intel is expected to retain a majority stake in Mobileye, an Israeli company that the chip giant bought in 2017 for $15.3 billion, the news service said. In December, Reuters and other business publications said Mobileye, which makes a camera-based system for adaptive cruise control and lane change assistance, could be valued as highly as $50 billion when it goes public—although recent market swoons mean achieving that lofty valuation is now unlikely.
Waymo clears a key regulatory hurdle to edge closer to offering driverless taxis in San Francisco. The autonomous vehicle arm of Google-parent Alphabet has gotten permission from the California Public Utilities Commission to begin charging passengers for ride-hailing trips in its self-driving cars, TechCrunch reports. But a human safety operator still must be present, according to the permit the regulator granted. The company had been offering unpaid trips to a select group of "trusted testers" since August. The company has also been offering paid autonomous grocery delivery since November.
Amazon's robotaxi arm acquires strawberry-picking robotics firm Strio.AI. Zoox, a driverless car company owned by Amazon, bought out the young Boston-based agricultural robotics company, which was only founded in 2020 and quickly found success in pilot deployments on strawberry farms in Florida and California, according to TechCrunch. Ruijie He, Strion's co-founder and CEO, will become Zoox's director of perception, while several other senior engineers from the robotics firm are also joining.
EYE ON A.I. TALENT
Global consulting giant Accenture has named Andy Hickl its global capability lead, artificial intelligence, the company said. He was previously a principal group program manager in the Azure "strategic missions and technologies" team at Microsoft, where he led investments in new big data, machine learning, and A.I. capabilities.
Insurance and reinsurance broker William Towers Watson (WTW) has hired Pardeep Bassi in the new role of global position leader in data science for the company’s insurance consulting and technology business, according to trade publication Insurance Business Magazine. Prior to joining WTW, Bassi was chief data science officer at LV=General Insurance.
Correction, March 9: An earlier version of this newsletter's "Eye on A.I. Talent" section misspelled the last name of Accenture's new hire Andy Hickl.
EYE ON A.I. RESEARCH
Understanding when self-supervised learning algorithms will fail. Self-supervised learning, which involves training an A.I. system on large amounts of unlabeled data, is a hot area of machine learning, responsible for some of the most interesting recent leaps in natural language processing and image classification, as well as systems that learn both images and language representations simultaneously. But A.I. trained in this way is also prone to failing in unexpected and surprising ways, which can make it problematic to use such systems in high-impact or safety-critical domains.
Now, in a paper published on the research repository arxiv.org, a team of researchers from Meta's A.I. research lab and the University of Maryland's College Park campus have proposed a way to predict when self-supervised systems will fail at an image classification task. The technique examines the different weights that a neural network applies to different parts of an image. It finds that in classes of objects that a system is likely to correctly identify, the A.I. has usually learned to put more emphasis on just a few key portions of the image that help differentiate that object from others. In classes of images that the system is likely to fail on, the weight the A.I. applies to different parts of the image is more homogenous, with no clear areas of focus, or the system tends to be weighting areas that don't seem to carry representational value, such as focusing on the background rather than the object itself.
Based on this finding, the researchers propose a "self-supervised representation quality score" or Q-score, that is able to predict, without any label information, whether a given sample is likely to be misclassified with up to 90% accuracy. It also found that the Q-score can be used to help regularize low-quality training samples, improving image classification accuracy of A.I. systems by 3.26%.
FORTUNE ON A.I.
Commentary: Your business needs an A.I. watchdog. Here’s how to make sure it has teeth—by Francois Candelon, Theodoros Evgeniou, and Maxime Courtaux
The cheap, slow, and bulky drones taking down Russian armored tanks for Ukraine—by Christiaan Hetzner
A.I. companies scramble to keep Ukrainian employees safe . There's a lot of A.I. talent in Ukraine, and Russia's invasion of the country has companies that have hired teams in Ukraine scrambling to get their employees out of harm's way. The Wall Street Journal has an excellent story looking at the efforts of one such company: New York-based Fractal Analytics, which had 81 employees, including a number of data scientists, in Ukraine before the war began, out of a global workforce of about 3,500. The newspaper chronicled the company's efforts, which began in the weeks leading up to the war, as the U.S. government and others sounded increasingly ominous warnings about Russia's intentions. The company hit upon a three-pronged approach that involved trying to provide employees temporary accommodation in Western Ukraine, close to the Polish border and away from where most of the fighting has been so far, helping to evacuate people out of Ukraine entirely, and providing others with extra cash. The company was able to assist 15 employees in leaving Ukraine, including at least one who chose to fly to Mumbai, where the company also has offices. About 40 staff are now in Western Ukraine, again some relocated with help from the company's travel department, but 20 remain in Kyiv, in increasingly dangerous circumstances.
From the story: As a CEO, Mr. [Srikanth] Velamakanni has regrets. Many Fractal employees in Ukraine didn’t think an invasion would happen, and Mr. Velamakanni said he now wishes he would have pressed more urgently for workers to leave Kyiv. The pandemic taught the 48-year-old Mr. Velamakanni that the contours of a CEO’s job have expanded to include looking after the well-being of an employee and their families, something the crisis in Ukraine only reinforced. “The boundaries of care, or your responsibilities, have just changed completely,” he said.
Our mission to make business better is fueled by readers like you. To enjoy unlimited access to our journalism, subscribe today.