Ethan Mollick’s One Useful Thing Substack newsletter is widely hailed as one of the best resources for understanding the rapidly evolving world of AI and LLMs. It’s no surprise that his 2024 book, Co-Intelligence: Living and Working with AI, is a thoughtful and widely practical guide for engaging with AI, regardless of the reader’s prior experience or technical knowledge.
Mollick’s approach is radically different than most other experts I’ve seen. With a background in management instead of computer science, Mollick is less concerned with prompt engineering and the technical limitations of various LLM models. Instead, he wants to find the best way to work with AI, and most importantly, to have AI support the best human work.
Principles for Co-Intelligence
Mollick introduces four practical principles for working with AI that will help anyone approach these new tools purposefully in the short and long term.
1. Always invite AI to the table
The capabilities of LLMs and other generative AI tools are evolving at a breakneck pace, and we’re constantly discovering new ways that these systems can be used. The best way to develop your knowledge of what AI can and can’t do and to find surprisingly novel use cases is to use AI in a wide variety of ways.
2. Be the human in the loop
At least for now, AIs can’t reliably work without human oversight. As a user, it’s important to stay engaged in the process to watch for hallucinations and strange outputs. AIs are tools, and like any tool, a responsible user is required for safe operation.
3. Treat AI like a person (but tell it what kind of person it is)
I think this is one of the most important insights that Mollick offers. LLMs work differently than most technology tools. Instead of highly structured code, LLMs are complex webs of training data and logic weights that produce strange emergent properties that we’re still struggling to understand. Strangely, the framing of how you prompt an AI can dramatically change the output, so providing gentle corrections and setting the right tone is an important part of getting the best results.
4. Assume this is the worst AI that you will ever user
This one is mindblowing and almost certainly true. Generative AI is still the newest trick at the party, and estimates are that nearly 70% of people have little to no experience with these tools yet. But they’re still evolving rapidly, and any limitation that you encounter while working with AI today will potentially be removed tomorrow.
Working with AI
Next, Mollich introduces several ways that AI can support humans in a wide range of applications. In the current loneliness epidemic, AI companions are already helping people mourn lost loved ones and may eventually stand in for friends or romantic partners. In other cases, AIs are helping brainstorm creative ideas, tutor both students and adults on a range of topics, and helping professionals workers write, code, and design. With so many use cases in the infancy of these techologies, we should expect that their impact will expand signficantly as the tools evolve.
Are we ready?
I think Mollich’s approach to AI is the best I’ve seen, and I’ve personally changed my approach to LLMs and other generative tools based on his advice. But I still have some real fears that things will get bumpier before they get better.
Near the end of the book, Mollich gives a few examples that jumped out to me. In one, two hypothetical workers are working in an architectural firm. One of the workers, Alex, stays up to date on the major trade resources and frequently asks senior designers at his firm to review his new designs. His coworker, Raj, incorporates AI into his work and use an LLM as a professional coach to reccomend interative improvements as he goes. In Mollich’s telling, Raj’s work matures faster than Alex because the LLM offers more specific and more frequent feedback as a sort of “ever present” mentor. And while this could one day be a possibility, I would counter that Alex’s face-time with the senior designers allows him to be seen putting in the work. I highly suspect that Alex would still advance faster in the firm than Raj, simply because Raj has been growing in private and Alex has been hobnobing with senior leaders.
A different sort of misalignment comes a few pages later. Mollich quotes the CEO of Turnitin, a plagiarism-checking company, who said “Most of our employees are engineers, and we have a few hundred of them….and I think in eighteen month, we will need twenty percent of them, and we can start hiring them out of high school rather than four-year colleges. Same for sales and marketing functions.” Mollich says that this reality is unlikely because most professional roles are much more complicated than we give them credit for. There may be mindless repetitive work, but there’s also a lot of judgement and decision making that AI isn’t ready for. But here again, I’m worried that Mollich may be wrong. Many corporate boards and executives are going to push AI too far too fast on the belief that workers can be replaced by AI models. I would be shocked if we don’t see a series of major corporate disruptions because LLMs were rushed as replacements to experienced workers.
In the end, Mollich posits three possible futures. Either AIs will improve no farther, they will improve slowly over many years, or they will improve massively very quickly. In each, the odds of disruption are signficant, and if you agree with McLuhan’s famous adage, “First we build the tools, then they build us,” we’re likely to see major shifts in human work, society, and identity.
Humans rarely appreciate significant moments in history as they’re happening, but the emergence of AI will likely be at least as transformative as the smartphone or internet (and potentially, much more so). Learning to live and work with these tools as soon as possible is the best way to ensure that AI helps us build a future that works for all us.