In my tenth column in the “Etzioni on AI” series, I want to share ten rules for implementing AI, preceded by timeless classics that still apply. They use my work with the founders of the AI2 Incubator, Madrona, and my experience as an AI inventor from Netbot (1996) to Vercept (2024).
To see what’s different about AI startups, let’s warm up to the ten commandments of startups in general, incorporating the wisdom of pioneers like Vinod Khosla, Reid Hoffman, and Eric Ries. Once you’ve mastered the classics, jump straight into the AI commands.
1. The real danger is regret. It sits in the stands at 60 and realizes you never stepped up to the plate.
2. Choose your founder carefully as a spouse. You may end up spending a lot of time with your founder in the early years. And the separation of founders is also very painful.
3. Increase your chances of success, not your identity. Don’t end up owning 99% of nothing. The right investor, incubator, or hire will add much more value than the cost.
4. Raise the story. Investors buy the company you can be. As Khosla puts it: “Don’t deviate your story from logical order.”
5. Solve the real problem. Is it a painkiller or a vitamin? The startup sells painkillers.
6. Talk to your customers early and often. Your guess is wrong until the customer proves otherwise; An ounce of data is worth a pound of insight. Scrappy testing involving users resolves conflicts that would otherwise take months to resolve.
7. Focus, focus, focus. The hardest word for some founders is “no.” Say it to good ideas so you can make one great. If your product is dessert topping and floor wax, then it is not.
8. Build a rocket while flying it. You start without all the answers in the area and iterate quickly to find more.
9. Be prepared to walk badly. Getting started can seem like a series of near-death experiences. The winners are the ones who refuse to give up.
10. Beware of consultants. Remember the ham-and-eggs adage: the chicken is involved, but the pig is committed. You want a dedicated team. Anyone using the meter has motives that are not aligned with yours.
Here are a few commands that didn’t cut it: hire carpenters, not architects; hire slow, burn fast; know your numbers are cold, especially your burn rate.
Other must-read ten rule lists include: Reid Hoffman’s, Howard Tullman’s, and Shlomo Kalish’s, as reviewed by Glilot Capital.
These classics still rule, but it’s time to add ten AI commands.
1. “We are an AI company” is not the same anymore. Table poles. Tell the story: what is the point of pain? Who is the customer? How do you make money? Why now?
2. Inadequate AI technology. As Madrona’s Matt McIlwain puts it, “the most important AI model is the business model.” And you have to deliver against that model – vision without execution is hallucination.
3. Do not put lipstick on the model. If your company is a deep glosser over someone else’s API, borderline labs will eat you alive. They have a model and distribution. You don’t have both. Build where they don’t (or can’t) go.
4. Own your data, but don’t mistake it for moat. Proprietary data is helpful, but it’s not a promised land. The key is the flywheel: Waymo learns from every mile its vehicles drive, and learning makes the next mile better. Create that loop.
5. Speed is the new ditch. AI has reduced the cost of intelligence. The edge is for anyone who is quick witted. As Bryan Kim of a16z writes, “momentum is the drain.”
6. Embed in workflow. The app that your customer opens at 9am and closes at 6pm is the one they can turn off. Be that. Or quietly take the one they are already using. AI coding makes it cheaper than ever to recreate an application layer from scratch.
7. Distribution is a scarce resource. Building an AI product has never been cheaper. Getting it in front of buyers has never been harder. A thousand companies competing for your customer will not lose features. They will lose where they are accessible. Distribution should be charged equal to the product from the first day.
8. Don’t marry a model. The frontier model that wins your demo today will be the third best in six months. Build a stack for easy changeovers when chemistry fades.
9. What is considered is the new COGS. Every question costs real money, and the cost is proportional to every user. Bessemer puts it bluntly: “If the math doesn’t work for 10 customers, it won’t work for 1,000.” Know how much your product is worth before you grow it.
10. Personal relationships are still very important. AI does not earn trust. It won’t hold your phone in the middle of the night or protect you from the board, and it won’t be there when you need a bridge circuit.
Seattle is an amazing place to live with these instructions: AI House is open on the waterfront; new funds have been raised; and hundreds of startups are thriving.
The time to start AI is now. Go build it.
