Summer into AI: Why Teams Are OK

I had an inbound question today that asked me about whether a team is OK in the competition. I responded, “Yep!” and then followed up with a “I’ll write a post later on about why you might not want to.” There’s good reasons to have a team-if you have someone younger in the family, for example. That makes a lot of sense to team up, just to give the other person a little guidance around what to do.

For the remainder of this article I’m going to outline why building out a broad team for an AI competition that focuses on shipping and proof is perhaps not an advantage. In fact, in many ways it might be a disadvantage.

The Coordination Tax

When you team up with someone I don’t view it as an unfair advantage for an ai-focused competition. In fact, it might be a liability.

You’ll pay for an overhead that doesn’t exist in a solo team: coordination and communication. Instead of rapidly iterating on and building out ideas, you’re having conversations-by the time the conversation is over, someone else in the competition will have already sent an AI to build two more projects.

This can be seen very clearly by way of example.

1. Open ChatGPT. Send a prompt to Chat to create an image.
2. While Chat is creating that image, hold a standard meeting with someone discussing more seriously what kind of image you should make.Review and identify what tool you should useReview who should review the output of the tool

If you’ve used AI at all, then you know option 1 is going to complete long before option 2 even gets added to the calendar.

Option 1 is so much faster than option 2 that it really upends a tremendous amount of organizational structure and theory. Dashboards, KPIs, communication and RACI start breaking down pretty fast in a fully AI powered world.

So much so that it just seems self-evident this is breaking a lot of pre-AI organizational design. If your constraint is how fast your development efforts are, then it would seem you already lost. You aren’t outcompeting the labs or companies that have banks of gpus they can toss frontier level intelligence at. It’s a simple calculation that you can test-run yourself: if your business depends on speed, how many open tabs of ChatGPT can you open?

Now imagine a company that does 100x that. You brought a horse to a formula1 race.

The Learning Opportunity

You could learn coordination approaches for a team with AI in this competition, I don’t disagree. However, this is a better way to build up yourself with new skills and capabilities. Turns out a lot has happened since January.

There’s a growing awareness that you can defer tasks and a certain amount of learning to AI agents to solve problems. You know what you can’t send out? Understanding. Can you be walking down the street with a keen insight about the project you built without actively participating in multiple levels of it’s construction? If you want to understand something, then you must do. A big reason this competition exists is to enhance the human brain to tip into understanding-expanding what’s possible, learn new tricks and capabilities, and even put to work some of the observations I’ve found around something as simple as writing.

That connects to the earlier note On letting AI do the writing.

It’s a solid read and why you should minimize how much AI does the writing, or find specific injection points where you re-write and/or just write something from scratch yourself. And in fact, I’ve even posted at least one writing exercise: Writing Exercise: Palimpsest.

The point here is related to understanding, but also brain science. If you defer understanding to the models, then you’re just going to have a bad time. In fact, did you know most models have a finite vocabulary - it’s interesting to me because you can see how a LLM tends to average out on certain words. This is especially true on AI heavy social media sites like Facebook or LinkedIn where the majority of posters are purely AI now.

The Context Window is You

AI gifts you leverage. A team can certainly take you further in some ways, deeper into a specific domain. For example, if I had an audio person to just focus on audio I could vastly scale out the number of sound effects and songs for the apps I’m slinging together. Except, now someone has to organize the work that needs done, bring it in, test and provide feedback and coaching. Alternatively, you can just direct the AI to get a “good enough” pass done and move on.

Work to be done is not a scarce input here. The scarcity is your time to drive AI tools to get something done. You have a finite amount of memory bandwidth. I have a lot of fun exercises you can do to find the upper bound of your memory without intentional practice (I’ll be adding my practice exercises to substack in the coming weeks). The thing here is that a team may not be the fastest path. If speed matters, which it does in a competition like this one-the faster path from zero to “shipped” is… you.

Thanks for reading! Subscribe for free to receive new posts and support my work.

Read the fine print in Summer Into AI: Are you ready to cook?.