No more painting by numbers
Every product builder I know is spending a significant amount of their time these days thinking about how to build with generative AI.
And, let’s not sugarcoat it…it’s hard to do. Even harder to do well. But there’s some value, I think, in describing how and why it’s hard to build with AI. So let me try to do that today.
The approaches I’ve seen to building with AI fit into four broad categories, each with their own challenges:
- Make the product experience better with AI
- Augment a human user with AI
- Outsource a specific task to AI
- Empower end users to build with AI
I’ll cover each below.
Make the product experience better with AI. I’d put features like LLM-powered summarizations (launched early by the likes of Notion as part of “Notion AI,” but now ubiquitous) or “chat with your help center” experiences into this category.
The obvious pitfall with these sorts of bolt-ons is when companies build “AI-powered” features that check the box (we launched an AI feature!) but don’t actually make the user experience any better. This is, unfortunately, all too common. Indeed, it’s amazing how many AI-powered features one comes across these days that essentially don’t work. Or only work sometimes. Which just isn’t good enough.
Augment a human user with AI. These are the “copilots” we’ve all heard so much about. It’s become a consensus view that developer copilots are incredibly useful (though probably not replacing software developers any time soon!). Unfortunately, I haven’t come across any other types of copilots, across any other domains, that come even close to the usefulness of coding copilots. (Please let me know if I’ve missed any!).
Model capabilities are partially to blame. (Foundation models are particularly good at coding, so it makes sense coding copilots would be the early breakout successes in this category). But I think the bigger challenge here is actually because of product design / UX. We just don’t have good product paradigms that I know of for a true “copilot” that sits alongside you, observes in the background, and only helps when it's useful. In early 2023, Diagram announced Genius, a “design copilot,” that would design right alongside a designer. As it turns out, it wasn’t real. But that demo remains probably the best example of what I imagine this could feel like.
Outsource tasks to AI. Generative AI models can be used to complete many tasks that have well-defined outcomes, but non-deterministic solutions (that is, there isn’t “one right” way to do it). That’s incredibly disruptive. (There’s a reason that the phrase “AI Employees” is on the lips of board members everywhere.) Examples of products leveraging this capability range from companies like Quilter or Robin AI (replacing a step in a process with AI) to Enso and 11x (replacing an employee with AI).
The challenge is that, despite best efforts, LLMs aren’t good enough at the multi-hop reasoning required to fully outsource complex tasks. This suggests to me that you’re better off narrowing down your focus so that, as Linus suggests below, you’re able to build a deterministic logic engine to guide the LLMs to do your bidding in a somewhat predictable way.
So, if you’re a product builder (or investor) targeting this space, and you want to find a sufficiently constrained vertical/function to target, where should you focus? I like Bill Gurley’s framework from the latest BG2 Podcast, where he argued that AI will win at “finite games” but struggle at “infinite games.” In other words, start by finding a function that is governed by “finite games” and build there.
Empower users to build with AI. This is what you see open-ended, customer-built, products like Zapier, Airtable, Notion, and even Adobe, Figma and Canva, attempting to do (as well as countless new “AI-first” startups hoping to disrupt these incumbents). And this is, perhaps, the hardest approach to get right.
Why? Because users don’t yet have an intuition for what AI can do. In this recent interview on the No Priors podcast, Howie Liu (Airtable CEO) shared that, especially amongst enterprise customers, users just don’t understand what these models are capable of. As Howie says, “even though there are so many different applications and you can apply AI to almost any use case in any industry, the gap right now is in imagination and know-how.”
In other words, if you want to succeed as a meta-platform that empowers users to build with AI, you first need to educate your users about what AI can actually do. That’s a tall order. Not impossible, of course. We’ve done this before when products with brand-new capabilities entered the scene (remember those old Macintosh ads…). But it will be a unique challenge these meta-platforms face.
–
So, it’s hard to build with AI. That much is clear. But what makes it so hard?
I think it’s because, as Sam Altman said, generative AI is kind of like an alien species. And it breaks most of our mental models for what it enables a piece of software to do.
Here’s what I mean. For the past decade or more, it was assumed that whatever you wanted to build was possible. There was basically zero “technology risk.” But it’s different with AI. You can’t just assume it will work (as builders attempting to outsource tasks with AI have seen). And even if it does work, it probably won’t work consistently enough (as builders launching AI-powered features that don’t live up to the hype know all too well). As a result, building with this technology requires a huge shift in the mindset of product and engineering teams.
And it’s not just that. As detailed above, you also can’t rely on well understood design paradigms to bring your product to life (as builders creating copilots have experienced). And you can’t assume your users understand the capabilities well enough to take full advantage of their power, either (as builders of customer-built products have seen). At each step, product builders are faced with another problem to solve. Nothing can be taken for granted.
That’s because what we’re really seeing, in my estimation, is the end of the “SaaS playbook” itself. All the foundations of product building that founders have relied upon for the past decade are being upended. And that makes everything harder. But, as I am nothing if not an optimist, it also suggests to me that many incumbents may have a harder time evolving than first imagined.
In other words, for the past decade, we’ve all been painting by numbers in the same SaaS coloring book. Startups and incumbents alike. Now, for the first time in a long time, we’ve got a blank sheet of paper in front of us. Game on.