There has been a lot of buzz surrounding the adoption of artificial intelligence. According to a recent report from McKinsey 57 percent of companies are now using AI in at least one function. But how much is hype and how much is built on a sound commercial base?
We spoke to Mike Loukides, VP of emerging tech content at O’Reilly Media and author of O’Reilly Media’s widely-cited AI Adoption in the Enterprise report, to discuss the current state of AI and what lies ahead.
BN: Are we moving beyond the adoption of AI because it’s new and cool to having a serious business case?
ML: We’re currently writing up the results from our own annual AI survey, and what we’re seeing disagrees pretty sharply with McKinsey. Roughly a quarter of the respondents to our survey have what we call a ‘mature’ AI practice, which means that they have one or more AI applications actually in production. That’s a lot less than 57 percent. More strikingly, our results have changed little since last year’s survey. Roughly half of the respondents were evaluating AI solutions — somewhat lower than last year’s numbers.
There’s obviously a lot of flexibility in what these terms mean and how respondents interpret questions. But it does suggest to me that, while there’s still a lot of experimentation going on, the market has stabilized. It’s less about hype, which has moved on to blockchains and NFTs. And getting past the hype will give AI the opportunity to prove its real value. There’s certainly been pushback against AI applications that have been adopted inadvisably — for example, resume screening. And companies are starting to understand what the real costs are: the costs of retraining models that have become stale, the cost of acquiring data, and the cost of integrating AI applications into automated deployment pipelines.
But I also think that businesses are starting to miss out. For example, we’ve seen tremendous advances in the ability of language models to develop code for programmers. AI will not replace humans; when used appropriately, it will help humans to be more effective.
BN: What, if anything, is holding back the widespread adoption of AI?
ML: In our 2021 AI adoption survey, we saw for the first time that the demand for expertise in data science (including AI and ML) was exceeding supply. So there’s a talent shortage. One thing that might help with that talent shortage will be the use of AutoML tools for building and training models. We are seeing some signs that those tools are used more widely in organizations that are newer to AI — as you’d expect.
There’s a bigger issue, though. For a few years, I’ve been saying that the elephant in the room is getting AI applications off the developer’s laptop and into production. This year, everyone’s saying, “Whoa, there’s an elephant in the room!” We need more people who know how to build data pipelines, test AI software, build deployment pipelines for AI applications, and all of that. It’s similar to what we call DevOps, or continuous deployment, or Agile, or something else. But AI applications aren’t the same as traditional web e-commerce applications. They throw a number of curve balls that don’t fit well with these somewhat older practices. For example, we know a lot about source management with GitHub. But with AI, you need similar management tools for training data. Those tools are only just appearing. We know a lot about testing. But how do you test applications whose behavior is statistical rather than deterministic?
So, in addition to the shortage of AI expertise, there’s a second shortage in expertise around deploying AI. Call this data engineering, ML engineering, or what you will; it’s a significant roadblock. And it requires new or improved tooling.
BN: Will we see an AI-as-a-service model becoming more commonplace?
ML: Yes. The cloud is an easy way to quickly assemble the computing power that you need to train models. AWS, Azure, Google Cloud, and IBM all have attractive tools for helping build and train models and these tools are particularly useful for organizations that are just starting with AI. We’re definitely seeing evidence that these tools are popular among those with less AI experience. What’s most interesting is that Microsoft Azure appears to have the lead, outperforming AWS.
BN: What role does AI have to play in the development process? How will developers need to adapt?
ML: I assume you’re asking about tools like GitHub’s Copilot and DeepMind’s AlphaCode. I think they will have a big impact, though maybe not the impact that people expect. From talking to people who have used Copilot in production, we’ve learned that it’s not particularly useful to new or inexperienced programmers. It’s not going to ‘steal programming jobs’ or anything like that. But it’s very good at making proficient programmers more productive. It lets them spend more time thinking about how to solve problems and less time looking up or trying to remember odd bits of documentation. It’s particularly useful for an experienced programmer who suddenly has to dip into a language that they’re less familiar with. And we’ve seen other AI tools to help programmers understand code that other people have written — reading code is a crucial, but undervalued, skill.
BN: Can we expect to see new job roles being created to serve the needs of AI?
ML: Definitely. Training and retraining models will become a specialization of its own, distinct from AI programming. Collecting and documenting data in appropriate ways will also become a new job role, along with other roles associated with data governance. Although many organizations haven’t realized it yet, we’re long past the time when you could use any data that was on hand, regardless of how it was collected, its impact on privacy, and biases inherent in the data itself or the data collection process. I’ve already mentioned AI operations, which will have to account for the differences between AI applications and the business applications we’ve become accustomed to running. While automation will play a part in all of these roles, humans will be needed to set the direction. I don’t think automation will eliminate any of these roles. Given the scale of modern data problems, I think automation will make these tasks possible.
Photo Credit: NicoElNino/Shutterstock