AWS VP talks up sci-fi maths tooling to rid world of software bugs

In a refreshingly frank conversation with media from around the world, a vice president at Amazon’s cloud arm poured ale-scented realism over the AI hype engulfing the tech world while bigging up the potential of an oft-overlooked branch of mathematics to eliminate coding blunders.

“Even the largest models are just machine learning. They’re not actually intelligent,” said Kurt Kufeld, AWS’s Vice President of identity, observability, and security when quizzed on the difference between AI and machine learning at the cloud giant’s annual security conference, re:Inforce, held in Philadelphia, PA., this year.

“They seem like it, but they’re not there… I don’t think we’re at true AI yet.”

Kufeld distinguished himself from the usual brand-peddling tech exec by repeatedly downplaying hyped claims of sentient software while making a compelling pitch for “automated reasoning” – using mathematics to effectively prove the correctness of code.

“It uses mathematics to basically try out every possible case so that you know for a fact that a certain piece of code works in a specific way and only in that way – hopefully the way you want it to work,” he explained.

Kufeld provided the example of the authorization runtime client – the critical code component that verifies a user’s credentials before granting or denying access to AWS’s services. His team used automated reasoning techniques to generate the code, then verify its correctness against formal rules.

“Not only did we prove its correctness, which by the way matters a lot because it is the single most important authorization piece of code probably in the world, but it’s also 53 percent faster,” he said.

The VP cited Intel’s notoriously buggy Pentium processors as a prime example of where such techniques could have prevented errors. One of Kufeld’s engineers formerly worked at Microsoft applying similar methods to tackle the “blue screen of death.”

However, he lamented that the community of skilled automated reasoning experts able to operate in a corporate environment is “extremely small” at an estimated population of under 2,000 worldwide.

On the AI front, Kufeld expressed concerns about the potential for large language models like ChatGPT to be abused for misinformation through manipulated videos and images – so-called deepfakes. He also warned of the dangers of miscreants using the code generation capabilities for nefarious purposes like surfacing vulnerabilities or crafting attack vectors.

The AWS exec dismissed the idea of an AI “kill switch” as fanciful, but advocated for regulations to govern the space: “We certainly need regulations around it… everyone plays by the same rules.”

While Kufeld may have punctured some of the more fantastical AI claims, his advocacy of automated reasoning injects some much-needed scientific rigour into the field of software development. Perhaps AWS’s commitment to the technology will inspire others – it could render entire classes of bugs extinct.

Read next: Cisco ramps up AI-era security with Hypershield