Antonym: The Cor! AI! Edition
An ad giant's AI strategy doesn't look good under the macroscope
Dear Reader
Questioning dodgy AI narratives is going to be a survival skill for investors, businesses, and employees this year.
I’m not talking about the basic BS-calls of the AI fearful – comments that starts with “it’s just..” is a tell – but the plausible narratives and polished PowerPoints of leaders and companies who are embracing AI without sufficient insight or humility.
The first principle of working with gen AI (and most complex fast-moving fields) is William Goldman’s eternal warning:
“Nobody knows anything.”
The second principle is Richard Feynman’s:
“The most important thing is not to fool yourself, and you are the easiest person to fool.”
If acknowledgement of uncertainty is omitted from a plan or prediction, if it is a fudged footnote to a pitch about how the next few years play out, it’s likely they are fooling you, and possibly themselves.
At best, they have fooled themselves in the telling of an exciting story, into making a possible future into a certainty. No one, including the storyteller knows if they will be right.
Publicis Groupe’s Core AI strategy
In an hour-long investor presentation last week, Publicis Groupe CEO Arthur Sadoun and a handful of senior executives laid out what seemed – initially – to be a credible, ambitious plan for how the advertising giant will build Core AI, a system that will “turn everyone into a data analyst” and give every one of its 98,000 employees access to insights and data on a massive scale.
It went down very well with the finance and trade press. For instance, Ad Age’s headline was:
Buyer beware: the problems are in plain sight if you look closely.
The investment doesn’t match the ambition: €100mn a year sounds like a lot of money but it is less than 1% of the group’s turnover.
Capability will hinge on learning: 50% of the investment will go into “people”. If that was just a training budget it would equate to c. €500 per person, the equivalent of a day-long course, or a subscription to a couple of online learning platforms.
Cobbler’s children effect: The strength of Publicis Groupe in AI is its digital transformation arm Publicis Sapient. But those people have day jobs and anyone who has worked on internal agency initiatives knows that your own ad, system, campaign, or project ends up at the bottom of the priority list after billable work.
In sections of the presentation executives showed what looked like demos of AI powered tools how the work of weeks can happen in minutes. One even said that they were showing the system in action with a made-up brand to protect confidentiality, but the real reason is that both the campaign and the tool were mock-ups.
The (very) small print on the slide with the timeline for development of Core AI shows that MVPs (minimal viable products) of elements of the system will be ready to demo by May and will be tested by small groups of users then “rolled out” in the latter half of 2024. Imagine what will have happened by then in generative AI: Google’s Gemini AI will be out with a version that rivals GPT 4, then GPT 5 is likely to be released, a model that will have at least ten times the complexity and power of its predecessor (estimated cost of training the model is $1.5 billion vs. the $100 million it cost to build GPT 4).
Publicis Groupe says the AI future is already here, it’s just not assembled yet. But it’s acting like it knows exactly how to do that. Its vision may stand as a cautionary tale to other companies looking to convince investors and clients that they have a handle on this whole AI thing.
Or they may nail it. Remember: “Nobody knows anything” – and that includes me.
Things I learned about generative AI this week.
Teaching AI is harder than it looks.
This week I got do to a couple of hands-on coaching sessions on working with generative AI. They were billed as prompting masterclasses, but I soon learned that teaching prompting is as complex and varied as the people who need to understand it. In planning the sessions I’d thought it would be all about practical skills and techniques: how to structure a prompt, how to build your own chatbot or GPT, and perhaps touch on more advanced techniques like “chain of thought prompts” and giving examples (known as “few shot”).
Start small and build your use of gen AI.
Two of the people I’ve coached in the last couple of weeks had the same issue – they wanted ChatGPT to help them with some of the more repetitive and dull aspects of research, so they could get into the exciting analysis and insights bit of the process faster. They were frustrated trying to find the right approach.
The blocker was that they were starting with trying to get a process to work in one prompt. In agile software development people use the metaphor of minimum viable product and then iterations toward a full product. You don’t start with a car, you start with a skateboard, then you make it more stable, then you add steering, an engine.
This isn’t a metaphor many of us have in our heads because we aren’t software developers. (Or we weren’t before ChatGPT.) With generative AI, we aren’t using a tool, we’re using a tool-making tool.
This is about “sharpening the saw”, doing the meta-work, knowing that 80% of the solution is asking the right question. The more we work with ChatGPT and similar models, the more it is clear that mastery comes from learning how to think about the work as much as how to use the technology. When we break down the steps of work like a piece of simple research into steps, into types of thinking we can see how each stage can be sped up or improved with AI. With each step we get better and see how to improve the whole process. The “final product” to step back into the agile software analogy, may simply be stringing those steps together in a prompt or a custom chatbot.
I’ll write more about this soon, no doubt. There may be a whole book in it.
Microsoft 365 with AI and Co-Pilot Pro are very powerful, but different.
Microsoft’s tools are powered by ChatGPT but very different to use in daily work. The Co-Pilot part of the business toolkit for instance, is amazing at helping you structure and think through things like project plans and proposals, but it is almost too keen to get started. I was halfway through giving it a brief and it just went ahead and wrote a whole plan. It was still super-useful as a starter.
Co-Pilot powered by GPT-4 is sometimes very intuitive and tells you how to do things before walking you through them. I’ve not got to grips with PowerPoint yet, but Word’s Co-Pilot features are great, as much for coaching and editing as for creating copy (I still prefer my own or bots that I’ve trained in my style).
That’s all for this week…
Thank you for reading – I hope there was something you found useful.
For more on AI, can I recommend Antonym’s sister newsletter, BN Edition from Brilliant Noise. This week we ran analysis on AI’s impact on Taylor Swift and deepfakes, the decline in search traffic to brands, and creative processes.
Antony