Moët Hennessy, Analog Devices and UC Davis build AI sensor for Champagne defect

Appian CEO proposes US AI alignment test, citing Hinton, Hassabis

Moët Hennessy, the luxury wine unit of LVMH, has collaborated with semiconductor and software firm Analog Devices and the University of California, Davis to build an AI-powered olfactory sensor designed to detect the Fresh Mushroom Aroma defect in Champagne grape juice before fermentation. In early testing, the system identified the defect with 99% accuracy, according to the companies and researchers involved.

The defect, undetectable to humans until after grapes are fermented into wine, has emerged intermittently in Moët Hennessy’s harvests and across the broader Champagne region in recent years, most recently affecting the 2023 crop. When it appears, the results can be disastrous and can cost the company millions of euros in wastage, said Manuel Reman, a member of the Moët Hennessy Executive Committee who oversees R&D.

Finding a way to detect the defect before the lengthy fermentation process was the focus of the collaboration. Analog Devices built a machine that used a new type of chemical sensor developed in-house to capture olfactory data from the grape juice. Together with Moët Hennessy and UC Davis, the firm trained an AI model to determine the likelihood that a sample was infected with Fresh Mushroom Aroma.

“I’m not telling you everybody was crying in the room, but nearly. I still have goosebumps. It’s so huge,” Reman said.

Max Shulaker, chief of Health Solutions at Analog Devices, said current sensing technology struggles with smell partly because it tries to identify and measure individual chemicals, a challenging task. He said he took a different approach, building miniaturized sensors that capture the overall aroma fingerprint of a sample. The AI learns to recognize patterns associated with outcomes of interest.

“Rather than asking ‘What chemicals are present?,’ we ask ‘Does this smell like a good sample or a bad sample?’ For many real-world applications, that’s the more relevant question,” Shulaker said. Moët Hennessy provided data from previous spoiled crops to help train the AI algorithm on what to recognize.

The system is not yet ready for production. The amount of time it takes to read a sample has come down significantly but remains around two hours — too long given the volume a Champagne house would need to run. A small house like Krug would need to run about 300 samples, Reman said, while a larger operation like Moët & Chandon might need thousands. Shulaker said he is confident about bringing that time down as the AI model continues learning from the samples it takes in, getting smarter about exactly what it is looking for. Reman said he hopes to put some of the devices in limited production next year.

Ben Montpetit, professor and chair of the Department of Viticulture and Enology at UC Davis, said the possibilities extend well beyond Fresh Mushroom Aroma. Smoke from wildfires is another problem impacting wine production that is often undetectable early in the production processes.

“In 2020 this cost the wine industry close to $4 billion in losses here in California,” Montpetit said. The collaborators are also exploring plant viruses that are emerging.

In a separate item in the same WSJ Morning Download, Matt Calkins, the chief executive of AI automation software company Appian, told a small group of reporters on Tuesday night that the United States needs to institute a strict model “alignment” test that would ultimately slow the pace of AI development.

Model alignment is a term commonly used by researchers to describe AI that acts in ways that match human intentions, and model misalignment refers to AI that acts in ways that ignore or conflict with human intentions. The idea has become more widely known recently as debate has grown over whether AI could one day wipe out humans. The WSJ reported that, at the extreme, misaligned AI models could see killing humans simply as a necessary step toward accomplishing their goals.

Calkins, who said he does not like to use “cataclysmic language,” still sees misaligned AI as a serious threat. “You come up with a technology as powerful as AI, with the capabilities that it has, and it’s just inevitable that 10 years from now, it’s either massively empowering us or substantially oppressing us,” he said.

His proposed solution is a model alignment test designed by top AI researchers and thinkers, including Nobel laureate Geoffrey Hinton and Alphabet chief scientist Demis Hassabis. “They would be willing to set the precedent for an alignment test that would be applied to the U.S. and would therefore be copied elsewhere,” Calkins said.

Calkins’s thinking is similar to that of leaders at Anthropic and OpenAI, who have said they are researching how they will make sure superintelligent AI models remain aligned. Both companies have said they do not yet have a reliable way to do so. Calkins doesn’t think the AI labs will go far enough on their own. “What they’re doing right now is testing alignment gently, realizing that their models are not aligned, and letting them loose anyway,” he said.

Appian is not strictly an AI company — the firm builds automation software that it sells to governments and enterprises. Calkins said talking about AI safety is not a matter of selling more software. “I was on an investor tour, and my CFO said, ‘Stop talking about alignment. This isn’t getting you any more investors,’” he said. “But I feel a duty to say this.”