The juice: where a knowledge worker starts with AI

Rob Dixon

Dec 23, 2023

5 min read

Manchester Metropolitan University asked me to do an evening Met Talks webinar on AI, and I set myself one question to answer: if you're a knowledge worker rather than a technologist, where do you start? Not a tour of the technology for its own sake, but a starting point for lecturers, marketers, finance directors, lawyers, and anyone else whose job is mostly thinking and writing. The recording is at the bottom. This is the argument.

Why it feels overwhelming

Sam Altman said that if you think you understand the impact of AI, you don't. I find that oddly comforting, because the deeper I go into this the more unknowns appear. It's the gift that keeps on giving. So if you feel overwhelmed, you're not behind. You're in the same place as the people building it.

Here's the bit that matters. In 1991 we had phone directories, paper maps, and magazines. My son is eight. He doesn't know what a phone directory is, and he thinks paper maps are what pirates use. The internet took thirty years to do that, and it took thirty years because it had to wait for everyone to buy a modem and for the newspapers to explain it. Generative AI needs no hardware and no journalist. It's on the phone in your pocket already, and it spreads at the speed of a tweet. Take the internet's thirty years and squash them into three to six. That's the window we're in.

Software has left the IT department

Before 30 November 2022, tech workers made the software. They built generic applications, we bought them, and they never quite did what we wanted. Since then, a knowledge worker can describe what they want in plain English and get it, and with a couple of no-code tools they can wire it into everything else they use. Coders got the same gift: write two pages and the assistant offers the next five. Two days' work in two hours.

That gives us two curves moving at once. The cost of making software is collapsing, and the productivity of everyone who works with words and numbers is about to jump. When Goldman Sachs says 300 million jobs are exposed, I don't read that as 300 million losses. Every industrial revolution so far has created new work and grown the economy. I read it as a very large productivity boost per person, and the question is who gets it.

The juice

Here's the framework I use for everything now. Knowledge work used to look like this: input, then we did our stuff, then output. Take a financial audit. The information comes in, we read the legislation, spot the anomalies, write the report, and send it out.

Now the middle of that is what the AI does. Looking things up, finding the anomalies, drafting the report, producing the one-page summary for the board, and turning the same material into a short training video for the manager who keeps miscoding things. All of that is production, and production is what the machine is for.

What's left for the human is what I call the juice. Clarifying what we're trying to do. Qualifying the input: is this the right information? Qualifying the output: is this right, and would I put my name to it? And defining the narrative. That's the job now. It's a smaller list than it used to be, and every item on it needs a human.

One more thing about output. Once the material is in, you can flip the format instantly: a book or a one-pager, academic register or the language of a twelve-year-old, English or any other language, a document or a video. Deciding which of those the audience needs is juice too.

There is no data in the model

If you take one technical idea away, take this. Google is retrieval AI. It indexes the web and fetches the ten most relevant pages, and it gives you the same answer every time. Generative AI doesn't look anything up. There is no data in it. It has been trained on an enormous amount of text and compressed into a model that predicts what comes next.

Try it. "For your birthday I baked you a..." Everyone says cake. So does the model. Now ask for a 3,000-word strategic analysis of Brexit's effect on the north-west economy. You'd mumble. The model produces it as quickly as it produced "cake", because it's the same predictive process. That's why it can create, and it's also why it can invent. I showed everyone on the call a perfectly good literature review on permitted development in areas of outstanding natural beauty, and then an equally confident one on sheep farming in central London among the blind community, complete with a journal that doesn't exist. Manage it the way you'd manage a bright new assistant: learn when to trust it and when to check before you sign.

The five things that make up AI literacy

Generative versus retrieval, so you know what you're holding.

Models, plural. ChatGPT is one of thousands, and there isn't going to be a single winner.

Prompt engineering, including saving the prompts that work so you can go back to the same consultant tomorrow.

Bots, which are just a model plus your prompts plus your documents, deployed.

Applications, broad ones like the AI in your office suite and narrow ones built for a single task.

And running through all of it, the ethics of what you put in.

Start with literacy, not policy

In the questions afterwards someone asked whether organisations should have an AI policy. Yes, eventually, and make it as light as you can: encourage it, play with it, and tell people what not to put in it. Never ban it. But notice the pattern. Universities leap to policy and ethics because that's what universities love. Businesses leap to innovation because that's what businesses love. Both are skipping the middle, which is literacy, and both end up writing rules or roadmaps for something they don't yet understand. Get everyone in a room, let them play, collect what they'd do with it in their own roles, and write the policy a few weeks later from a position of knowledge.

And I'll criticise my own session while I'm at it. Don't learn this by going to talks or reading books. Learn it by doing. Lots of little goes, one real task at a time.