The robots are coming

If you think your working life changed a lot in the past three years, buckle up: It’s going to change even more in the next three.

That’s because artificial intelligence is now mature enough to dramatically change the way many of us do our work. And while it may be possible to dodge AI’s impact for the next three years, embracing it now is the best way of nurturing your career prospects for the longer run.

 

GPT will change our work

While AI has been around for a while, what’s changed in the past year is what’s known as “generative AI”: tools like ChatGPT (which can write text or code) and DALL-E (which makes images). You may find their creations unconvincing, frivolous or even unethical, but I don’t recommend dismissing them too quickly.

GPT-like tools are going to make their way into most corners of the working world over the next few years, in ways that will transform or eliminate a lot of jobs (as well as creating new kinds of employment).  The best way to be on the winning side of that equation, and to play a role in shaping the business and social impact of the AI transformation, is by understanding the way these technologies may affect your own field. That means rolling up your sleeves, and learning how to use them.

I expect that in my own working life, GPT will significantly change my opportunities as a writer and data journalist; only my speaking feels likely to continue in something like its current form. That means l’m looking for all the ways AI can extend or transform the value of my writing and data work. My strategy: Commit to regular experiments that build AI fluency.

The technical details and how-tos on these experiments are probably interesting only to fellow geeks (geeks, feel free to hit me up for details!), but let me offer some recommendations that apply to everyone.

 

Give all your work to the AI 

A good starter experiment is to try giving your work to ChatGPT or another AI. If you want to know how much of your work is potentially subject to automation, the best way to find out is by continually trying to automate it yourself.

Don’t just give ChatGPT a prompt like “Write an email to a potential client asking to book a meeting”; first provide a couple of examples of your own past emails to work from, and then give your request as a prompt. This is called “few-shot learning”, and since the format matters, it’s worth looking at how to structure your examples. (This post helped me.) You’ll need to give ChatGPT a few prompt/reply pairs before leaving your final prompt open for ChatGPT to draft a reply.

By giving ChatGPT some examples of your own work, you increase the odds of it doing at least a decent first draft. Use this approach for various kinds of work, and try each one again from time to time (particularly as new GPT models come out) so you can see how the model is improving.

 

Include AI in your workflow 

Even if ChatGPT or other AI tools can’t yet do your tasks for you, they may be helpful in your own work process. I recently gave ChatGPT a document I needed to re-draft, but was avoiding; I asked it to re-write my copy, aimed at a specific audience. Its draft was not that great, but it got me going on my own work and helped me do my re-write a lot more quickly (and less painfully).

Look at the places where you get stuck in your work, where you have repetitive tasks or where you find yourself routinely referring to documentation in order to get your work done. These are the kinds of tasks that you can expedite with AI. Some examples:

• Use GPT for Sheets & Docs to bring GPT right into Google Sheets (or Docs) where it can help you write formulas, taglines or other texts (my colleague Ian Hensley put me onto this handy tool!)
• Try out one of the examples in the OpenAI playground to see if you can automate a task like creating a spreadsheet or classifying data.
• Turn an outline into a first draft by pasting an outline into ChatGPT and asking it to draft a document in a specific format or tone, based on that outline.

Hone your AI skepticism

Just as important as learning what you can do by including AI in your workflow is learning what you can’t. I used the OpenAI playground’s spreadsheet creator tool to generate a list of LinkedIn posts and URLs about remote work…but when I clicked on the URLs, I discovered they were fictional!

Understanding GPT’s propensity for hallucination is crucial to developing your own AI bullshit detector. Generative AI is going to drive an explosion in fake content—folks, the years ahead are going to make Putin’s fake news machine look positively tame!—so we all need to continually hone our internal Turing tests.

 

Structure your output

The next step up from few-shot learning is fine-tuning your own GPT model: building a dataset of prompt/completion pairs so that you can use GPT to draft text, answer questions or solve problems in a way that is specific to your particular needs.

This is not a quick or easy process, but it’s more accessible than I’d anticipated. Even though I’m not a programmer I was able to follow the OpenAI documentation and get a fine-tuned model up and running. (The process was a lot faster because I did my monkeying around with an actual programmer sitting beside me to answer occasional questions.) It took me all of an afternoon to build a chat model that generates imaginary anecdotes about my kids, based on the contents of a spreadsheet of cute kid anecdotes I’d exported from Facebook and lightly cleaned and categorized.

Most people won’t regard custom-training a GPT model as a fun way to spend an afternoon, and most people don’t have spreadsheets of categorized kid anecdotes lying around. But everyone can start to structure and categorize their own work for the purpose of future training.

Right now, you need to be a bit of a geek to turn your spreadsheet into a custom GPT model, but I promise it won’t be long before that process becomes easy and accessible. (Though not necessarily cheap: Converting my 1,200 kid anecdotes into a GPT model cost $10 in processing time, and my next experiment cost me $25 in processing.)

I’m a spreadsheet lover and a chronic categorizer, so I already have many portions of my work organized in spreadsheets or in structures that can easily become spreadsheets. The more consistently you structure and categorize your work, the easier it will be for you to turn into training models later.

Some ideas:

• Pick a type of email you need to address routinely, and start labeling or filling prototypical examples with a single folder/label. You can use these messages (and your replies) in a future training model.
• Start using a spreadsheet, database or table-based app (like Coda) for some portion of your work that requires managing information or knowledge. For example, I keep all my story ideas in a single Coda table; I’m planning to take my past ideas and turn them into a training model.
• Think about where you have information stored in ways that you can export or convert to CSV. For example, you could export or scrape your social media posts to a CSV file that you could use as the basis for a future training model.

Engaging with our AI future

If you’re wondering why you would want to make it easier for your employer (or someone else) to automate your work, it’s time for a reality check: If something can get automated, it will get automated. You get to choose whether you’re driving, managing and adding value to the automation process—that is, making yourself a valuable enabler and multiplier of the value that’s created by generative AI—but you’re probably not in a position to stop it.

This may be the moment when some people object: But we have to stop it. We’ll find some way to regulate, bottleneck, slow or prevent the employment-busting impacts of AI, because we have to.

But we have to is magical thinking. We have to becomes reality—in the form of policy change, the creation of new jobs or the development of new technologies that channel AI in a different direction—only through the actions of citizens and leaders who deeply understand the technology we have, and the road we’re on.

You can be one of those citizens; you can be among those leaders. But that requires the kind of understanding you can only develop by making AI part of your working life, beginning today.

 

Covid was our fire drill

Maybe it seems like a tall order to transform your working life this quickly; to become literate in AI, and in a whole new way of working, in just a year or two.

But let me remind you: You’ve already done it.

When Covid hit, most of us had never worked remotely; many of us had never done a Zoom call; plenty of organizations still had team members who resisted using Slack or Google Docs.

Now these ways of working are woven into our daily lives, and while we’re still figuring out how to make this new model collaborative and sustainable, few of us imagine it’s going away.

Covid was our fire drill. We changed our way of working, and we changed quickly.

Now we need to change again, because the robots are coming.

Time to roll out the welcome mat.

This post was originally featured in the Thrive at Work newsletter. Subscribe here to be the first to receive updates and insights on the new workplace.