What Cheetos tell us about enterprise AI

The Big Story

Cheetos and AI

Enterprise AI is in a kind of weird limbo. It's not to say AI isn't powerful; it very much is. But the fear of woeful returns on investment has shaped how companies are rolling it out.

• Research has shown that organizations are opting to keep AI projects in the pilot stages, segmented to individual business units or specific tasks, while they determine what, if any, return on investment there is.
• Most experts would say the true benefits won't be realized until the projects are deployed at a much larger scale, but that is an immensely difficult — and costly — task, which involves making difficult organizational and cultural changes for the AI to be successful.
• Still, the pandemic definitely accelerated a lot of the ongoing AI efforts: If companies weren't already exploring what can be done to make stored data more than just a landfill, they definitely are now.
• And for those organizations that had already embraced AI, ongoing pilot projects are now nearing the critical point of proving value or becoming yet another cautionary tale.

Cheetos provide a useful window into enterprise AI adoption and what it might really look like for many companies. No, really, hear me out. Since early 2019, PepsiCo has been working with Microsoft to use AI to try to make the, uh, perfect Cheeto. And Denise Lefebvre, the senior vice president of research and development for global foods at PepsiCo, swears it is more complicated than it sounds.

• To make Cheetos, the company uses what's called an extruder: basically a machine that turns Cheeto mixture into Cheetos that are ready to be cooked. The results can vary depending on factors such as water temperature, ingredient levels and the operation of the extruder itself.
• Those factors and processes are currently set by human operators, who have to simultaneously monitor many different metrics to know what to change.
• "It's a quite highly-engineered product," she told Protocol. "Cornmeal variation, steam pressure variation, water variation…."
• With Microsoft, PepsiCo is hoping to automate aspects of that production to allow for a more consistent quality, while simultaneously reducing waste in the process, per Lefebvre.

But a perfect Cheeto is a lot of work for machine learning. So alongside experts from Microsoft, PepsiCo sought input from chemical engineers, food chemists, sensory scientists and the machine operators who actually have to work alongside the technology. (Microsoft calls this feedback loop "machine teaching.")

• PepsiCo also made use of advanced simulation technology, a technique that companies are increasingly turning to to help train the AI systems, to "simulate an entire day's run in the span of 30 seconds," said Microsoft's Mark Hammond, the general manager of its autonomous systems unit. There are limits to the simulations, but "you don't want to do your initial training and testing" on live machinery, Hammond added.
• This is where the problem differs from, say, AI that helps with recruitment: These aren't major data sets that the system is analyzing. It's automating a very manufacturing-heavy process based on countless simulated test cases.

Slowly, PepsiCo is putting this AI to real use. After testing the system out in a pilot plant in the U.S., the company is putting the technology into actual production in one of its facilities in Spain. And it's already testing out how the system could be used in the production of Doritos.

• For Hammond and Microsoft, this is an example of how to move AI out of the research phase into actual real-world use cases.
• There's "tremendous research that gets published and shared widely and gets a lot of hype associated with it, but the direct ties of that research to the applications in industry are not always readily apparent," he said. "Our focus is quite squarely on those [real-world] applications," he added.

While solving for the perfect Cheetos seems trivial, projects like this reveal — and help to solve — the operational challenges that are preventing more enterprises from taking their own AI projects into large-scale production. The more companies that try to solve their own Cheeto problem, the sooner enterprise AI inches out of its limbo and into the mainstream.

-- Joe Williams

A MESSAGE FROM INTEL AND MICROSOFT AZURE

For corporate IT managers, there are many motivations to move dynamic workloads to the cloud. It provides an irresistible trifecta of flexibility, scalability, and costs savings for those managing varying workloads. Here's how to keep your data safe while it's in the cloud.

Read more.

This Week On Protocol

Digitizing Wall Street: Investors — well, at least, non-Reddit investors — are among the most avid users of data, deploying advanced analytics to help pinpoint trades by the millisecond. Now, Snowflake is teaming up with marquee institutional investor BlackRock to help customers make even better use of stored information.

Turmoil in Google's AI team: Well, more of it.The tech giant told employees it will change some of its hiring and firing policies in the wake of continued firings in its ethical AI team, but Timnit Gebru was less than impressed by the outcome, reports Anna Kramer. This whole saga has been a huge mess for Google and it seems very eager to move past it, installing long-time engineering head Marian Croak as the new leader for the responsible AI team.

Doubling down on D&I? Holler CEO Travis Montaque has sound advice on corporates who want to do that. Of note, he argues that tech companies should approach the issue in the same way they approach engineering challenges: tapping the agile method.

Coming up this week

Lots of people will be eager to hear what Nvidia has to say about the chip shortage during its investor call. Attention will also be on two other earnings calls this week: Salesforce (expect questions about Slack) and VMware, which is still seeking a permanent CEO.

Feb. 23: The Startup Grind conference begins with speakers including Slack Chief Product Officer Tamar Yehoshua.

Feb. 24: Nvidia reports earnings. Startup Grind continues with sessions featuring execs from Google Cloud, Twilio, Snowflake and more.

Around the Cloud

• Mailchimp is still reeling from the allegations of sexism and bullying by a former senior manager.
• IBM may be trying to sell IBM Watson, one of its flagship AI offerings. It's yet another sign that nothing is safe under Arvind Krishna as he remakes the struggling tech giant.
• LinkedIn is gearing up to enter the freelance market, The Information reports. That apparently looks like a booking service akin to Upwork or Fiverr, planned to launch later this year.
• Keeping on the AI-for-food theme, Wendy's is also testing out how AI can better predict what customers are going to order.
• Low-code provider Creatio raised $68 million, the first time it's solicited outside funding.
• A sad moment for true AI believers: Newly-pardoned Anthony Levandowski is closing The Way of the Future church.

A MESSAGE FROM INTEL AND MICROSOFT AZURE

For corporate IT managers, there are many motivations to move dynamic workloads to the cloud. It provides an irresistible trifecta of flexibility, scalability, and costs savings for those managing varying workloads. Here's how to keep your data safe while it's in the cloud.

Read more.