generative glitter #02
Labor, augmented efficiency, becoming a human guide
generative glitter is a weekly newsletter focused on how to use generative AI to work better—from theory to techniques. Let’s go!
First off, this week OpenAI announced some new data privacy features and an upcoming ChatGPT Business plan. This new plan is specifically designed “for professionals and enterprises who need more control over their data as well as enterprises seeking to manage their end users.”
Meanwhile, HuggingChat, an open-source alternative to ChatGPT, was just released, and it's also free to use. (Shout out to Adam Yeatts for bringing up the competitive space here.) It's hosted by HuggingFace, an AI community and platform, and it currently uses a model built by OpenAssistant on top of Meta's open-source LLaMA model. There's a lot to think about there! Maybe give it a whirl and see how it does.
Down to business, let's talk about the elephant in the room: automation and labor. There's been a lot written about the impact of AI (and automation more generally) on jobs. People are afraid that generative AI will replace them at work.
Since lots of people are talking about automating lawyers these days, let’s use the legal profession as an example. If it becomes very fast and cheap to produce legal documents, then the value of individual paralegal services will go down. It is easy to imagine a future where more firms use AI instead of people.
As an example of a labor group in another industry taking action to avoid this, the Writers Guild of America—a union representing Hollywood writers—has added an objective to "regulate use of material produced using artificial intelligence" in contract negotiations this year.
The editorial that I referenced last week points out that, instead of doing the same jobs with fewer lower-skilled people, businesses could just as easily let workers add more and new value with generative AI. So maybe the paralegals and the screenwriters can pivot to using new AI tools, which would either free up space for more of the same work (if there is latent demand) or for new work.
There are two relevant labor theories that Packy McCormick brings up in their Not Boring newsletter this week: Parkinson's law and Jevons paradox.
Parkinson's law says that if work takes less time to complete, then the amount of work increases to fill the space.
Jevons paradox says that falling costs for resources can increase demand.
They suggest that these effects could play out with products and services that are accelerated with AI.
So what does this mean for workers? Well, if Jevons paradox holds true, then there could be more work to be done, not less. And if Parkinson's law holds true, then we might fill our work time with other tasks. Whether those other tasks are better or not will vary by individual; what if you really enjoyed or were skilled at a task that was just devalued (like writing legal briefs)? That would suck.
All of this assumes that the employer is the buyer, though. With their ChatGPT Business plan, OpenAI is aiming to fill that niche. They’re betting that businesses will buy tooling for their employees and that they will want control over data and permissions (seems reasonable).
But, as long as tools like ChatGPT and HuggingChat are open and accessible, then individual workers could also use the means of production to their own advantage. Imagine the screenwriter using AI to write new material, instead of the studio replacing the screenwriter with an AI. That is what this newsletter is generally about.
What do you think? Any tasks at your work that you wish you could automate or augment? Reply and let me know!
Some evidence
Another research article is out from MIT this week showing that generative AI is a powerful assistant at work. The study included over 5,000 customer support agents in a real-world (non-lab) setting. The agents who used ChatGPT saw a 14% increase in issues resolved per hour, especially among inexperienced workers. They also found improved customer sentiment and employee retention in the experimental group.
Anecdotally, I’ve seen this first hand, from computer programming to data entry.
As an inexperienced javascript coder, I’ve found GPT really helps me get to some code that works faster (I’m looking forward to quantifying this in upcoming issues).
I’ve also seen faster and higher-quality data entry among colleagues who get a generative AI assist, and the impacts are definitely greatest for inexperienced workers who might have struggled with the task otherwise.
Don’t take my job
focus on becoming thinkers who guide the AI in executing - Josep Ferrer on Medium
Have you ever experienced GPT output that seems kind of mid? Well, that's kind of what these models do.
This reality makes generative AI awesome for tasks like writing code, drafting reports, or replying to customers—which tend to work better when they are predictable.
The same bug that limits originality and uniqueness for making artistic content can be a feature for doing consistent and reliable work.
Getting good results instead of boilerplate text, though, still requires some work. You need to guide it. Here's an example:
Write a short message for my team about how they did a great job completing the development sprint.
I just wanted to take a moment to congratulate all of you on completing the development sprint! You all did a fantastic job and should be very proud of the work you accomplished. Your hard work, dedication, and collaboration paid off, and the results speak for themselves. Your commitment to quality and teamwork is truly impressive, and it shows in the exceptional work that you've produced. Keep up the great work, and let's continue to build on this success in the future!
Not bad, but kind of generic. I’d want to basically rewrite it before shipping it. Let's give the model even more context:
Write a short Slack message for my team about how they did a great job completing the development sprint. Start with a general statement and end by highlighting the work to refactor the backend code, which boosted our system performance.
Hey team! Great work on completing the development sprint! I wanted to give a special shoutout to everyone who worked on refactoring the backend code - your efforts have significantly boosted our system's performance. Keep up the amazing work! 🎉🚀
Now we’re getting somewhere…
The point here is that your results depend on how much you create guardrails with the input. The job changes from choosing the words to choosing the concepts.
Want to explore how to apply AI in your work? Reply to set up a free discovery call with me. I want to hear about your ideas!