In this talk, Toju Duke draws on six years of AI experience, including work on Google's early generative models (LaMDA, Imagen, GLAM), to highlight the rapid rise of AI adoption and the common lack of deep understanding among practitioners. He stresses that while AI can appear deceptively effective, careless use leads to inefficiencies. Duke revisits the classic machine‑learning lifecycle—planning, data collection, feature engineering, model training, deployment, monitoring—and focuses on the first stage: planning and ideation. He argues that successful AI integration requires deliberately designing AI‑driven workflows rather than indiscriminately plugging AI into every task. Using marketers as an example, he explains how a workflow combines large language models and autonomous AI agents to execute a sequence of tasks, thereby preserving cognition, critical thinking, and agency in the user.
Chapters
After working in the field of AI for 6 years and being on the leading team that built Google's early generative AI models such as LaMDA, Imagen, and GLAM, I've observed the speedy rise of AI adoption and noticed that lots of people lack the depth of knowledge on how to use AI effectively or even how it works. Of course, we weren't taught AI in schools, and it's still very much an emerging technology, so no one expects a thorough grasp of its workings without prior education. But the thing is, AI appears to work very well, so well that it can be very deceptive, which can lead to work slop and inefficiencies. But it does work, and it's important to adopt AI and embrace it because it does actually help save you time and become more efficient. But there's a trick to this. In the very early days of machine learning development, basically AI, the ML lifecycle was always recommended in AI implementation and deployment. It involves planning, data collection, feature engineering, model training and architecture testing, deployment, and monitoring. The very first stage of the ML lifecycle is what I'll focus on, which is planning and ideation. When working with AI, you need to plan and understand the areas you need the AI to execute. For example, if you're a marketer drowning in 100 million tasks, you may decide to use AI for content creation. Yes, I know most marketers are already doing this, but, but here's the trick. It's not just about focusing on plugging AI to every single task you need to execute. It's about workflows. When you create an AI workflow, you would have decided beforehand the various tasks you need to achieve and understand what an AI workflow would look like to achieve your goals. AI workflows are basically a combination of LLMs and AI agents working autonomously to deliver the tasks been asked to do. Implementing a workflow protects your cognition, critical thinking, and agency, which is critical in today's world.
