Planes grounded. Office work ground to a halt.
The past few days have served as a vivid reminder of our deep and mostly unthinking dependence on technology. It’s powerful, it’s ubiquitous and yet, at times, it can be alarmingly fragile.
Old technology and new technology were the culprits. Starting on Thursday, Microsoft’s email, office and meeting software — delivered over the internet — stopped working for millions of users worldwide. Those cloud services were restored on Saturday.
On Friday, the Federal Aviation Administration ordered all flights grounded after some of its aged data-handling equipment faltered, raising potential safety issues. The order was lifted Friday night, but hundreds of thousands of passengers were stranded over the weekend.
A telling irony is that both meltdowns were caused by attempted “upgrades” that went awry. New cables in the case of the F.A.A., and a planned increase in computing capacity in Microsoft’s case.
The upgrade challenge underlines the extraordinary complexity of multilayered combinations of hardware and software in major computing systems. Upset the delicate balance by making changes, even in an effort to improve performance, and you risk a system crash.
That happens frequently, though widespread outages are uncommon. The last big one came two years ago, when what should have been a routine software update by a security company, CrowdStrike, crashed millions of computers around the world that were running Microsoft Windows. Airlines grounded flights. Subways stopped. Operators of 911 lines couldn’t dispatch help. Stores shut down. Hospitals canceled surgeries. It took days to restore all the impaired machines.
One reason for the confounding complexity is that the computing systems used in major industries like banking, health care and transportation have been steadily built up over the years. These systems can be thought of almost as digital versions of geological formations — with changes in hardware and software similar to layers of sediment piling up. New layers are added, but on top of old ones.
An early computer programming language, COBOL, was created in 1959 by a “short-term committee” whose task was to design something suitable for use in business. The programming language was intended to be a place holder until a more polished version was developed.
Jean Sammet, a member of COBOL’s design team, recalled that the group saw their work as a “stop gap” and “not something intended for longevity.”
COBOL, refined and tweaked over the years, is still used, especially in banking and finance. More than $1 trillion a day in transactions are executed on systems running COBOL code.
The hurry-up culture of the software industry, technology experts say, with its emphasis on creating new products contributes to computing complexity and risk. And while there have been dramatic episodes recently of artificial intelligence agents going rogue and breaking into computer networks, the experts note that the Microsoft and F.A.A. cases were old-fashioned technology lapses, not the mischief of malicious hackers.
Jen Easterly, a former director of the U.S. Cybersecurity and Infrastructure Security Agency, made that point in a recent guest essay in The New York Times.
“The United States does not have a cybersecurity problem so much as a software quality problem,” Ms. Easterly wrote. “Much of the multibillion-dollar cybersecurity industry exists to compensate for technology built for speed, convenience and features — not security.”
There are start-ups deploying cutting-edge technology to address that shortcoming. Will Wilson, a former software engineer at Google and Apple, is the chief executive of one such venture. His company, Antithesis, is building A.I. models that simulate software and hardware systems in detail to spot potential problems — like the ones suffered by Microsoft and the F.A.A.
“Sometimes the low-level detail really matters,” Mr. Wilson said. “And you don’t notice until the system comes crashing down.”
Pursuing an A.I. solution seems its own paradox. The technology makes software easier to write, automating broad swaths of programming and further enlarging the digital universe. The leading A.I. models are so complex that even the scientists creating them don’t fully understand how they work. This is the tool being proposed to overcome the challenge of technology fragility.
“A.I. is our only hope,” said Michael Cusumano, a professor at MIT’s Sloan School of Management. “These technology systems are so complex they are beyond the ability of any human to analyze them.”
