In her keynote at #mtpcon London 2023, Claire Woodcock, Director of Product, ML, at Mozilla explains how to keep your head about generative AI, when everyone else loses theirs. Watch the video in full, or read on for highlights from her talk.
Key takeaways
- Powerful human stories have made AI conversations mainstream
- AI seems so human and simple to use
- Technical leaders are under pressure to implement AI products, maintain technical advantage and leverage perceived productivity and profitability gains
- Using generative AI without proper testing or consideration can lead to dangerous outcomes
- The market is flooded with generative AI options, which means everyone can do the same as you
- Leverage product fundamentals when introducing AI into your product
Claire begins by looking at the hype around generative AI, and how this technology is not new. She first “realised AI was part of the future” when working in game development, before she moved to commercial AI labs and machine learning-powered products. Recently foundation models, the models which power OpenAI, have been attracting attention, though currently, 80% of product people aren’t confident about how to use AI, Claire says.
What is generative AI?
Claire outlines the basics, exploring how we describe a machine-learning model that can create content on its own (text, image, audio), from representations computed from vast amounts of data when we talk about generative AI. It’s something that ‘“your average startup can’t afford to build”, she says. Models are tuned and prompted to perform tasks and answer questions. Claire says that while this technology has been in design for years, we’re now at a “magical moment where things seem to have clicked into place”.
AI today feels surprisingly creative and effective, Claire says. She highlights an image from Midjourney that “looks pretty good” and a news story about a sick child who was diagnosed through AI and successfully cured. Powerful human stories have made AI conversations mainstream. Because AI seems so human and simple to use, technical leaders are under pressure to implement AI products, maintain technical advantage and leverage perceived productivity and profitability gains.
Deciding when and how to use generative AI
Claire shows how companies using generative AI without proper testing or consideration can have dangerous outcomes, for example news sites generated headlines which described a dead sportsman as “useless”, and a charity’s eating disorder helpline gave calorie-reduction advice to users. “The basics remain the same,” Claire says, “we need to build products that serve our user’s needs,” and importantly test before taking something live. The market is flooded with generative AI options, which means that everyone can do what you’re doing, so how do you decide where to play?
Claire recommends the Kano model in these circumstances. It plots customer satisfaction against feature satisfaction. It’s important therefore to find something unique, defensible and useful to build a delightful differentiator, she says.
“If generative AI isn’t your core value proposition, you want to try it in lower visibility areas first,” Claire says. We can always learn how AI is used elsewhere before building with it. Claire compares uses of AI in industry and how performance and interaction is already different for adults and children. For example, Salesforce AI has a conversational interface, while Roblox, a video game, is heavily prompt-driven and less conversational. Claire says: “Children have not been primed with two decades of mobile phone usage like we have, so this gives you a more strong indicator of how people who haven’t been primed by older technology are going to respond.”
Five steps to get generative AI into your product
In the last part of her talk Claire introduces five steps to get AI into your product, leveraging product fundamentals, as well as machine learning and generative AI specific items:
- Select your problem: Look for something low profile, high impact, an area for efficiency and not a part of your core customer-facing offering.
- Evaluate your data: As data is core to a successful machine learning solution, look at the quality of data you have, or how to collect, store and govern that data effectively.
- Decide on build vs buy: Tuning a model and managing infrastructure and servicing it yourself requires skills, but a third-party comes with a premium cost and requires a serious consideration of data privacy and IP leakage.
- Build a proof of concept: Test that quality meets production needs and that the data accuracy, tone and investment make sense.
- Scale your solution to production: Be conscious of spiralling costs and model drift, where behaviours change over time.
Claire closes by saying “the hype is pretty real, but the product basics stay the same […] focus on your data quality, and test and learn, and monitor that model once it’s in production”.
Want to turn the learning from this talk into action?
Our exciting #mtpcon London 2023 Keynote Kit brings you all of the insights from our London keynote talks plus additional helpful discussion points and thought starters so you can easily translate the insights into actions – perfect for making effective improvements to your role, team, and product. Plus, get an email notification when each talk is published! Sign up now.
