Key Takeaways

  • Forget detailed 9-month product roadmaps in AI. Andrew Ambrosino, product and engineering lead for the OpenAI Codex app, says any precision beyond a few months is "false precision" because model capabilities change too fast.
  • Ambrosino's strategy for AI product planning is to prototype all potential future features for the next year or two. Ship what's viable with current models, and let other ideas "sit and bake" until new model leaps make them feasible.
  • The success of the OpenAI Codex app, released in February, was entirely dependent on a few months of model improvements. Ambrosino says the app “would have absolutely failed in the market” if it had been ready in November.
  • Don't stubborn yourself into thinking a feature is bad just because it's not working. Ambrosino pushes teams to understand that it “might not be ready yet” for the current state of AI models, even if the vision is sound.

The Method

Traditional product roadmapping is a liability in AI. Where most teams meticulously plan out feature sets for quarters or even years, Ambrosino explains that for AI, “the shorter term something is, the more detail it needs. And then it's not that we don't plan for 9 months out, it's that that just has to stay very hazy because any amount of precision that you add to a 9-month plan right now is false precision.”

Instead of fixed timelines and detailed feature specs for the distant future, Ambrosino's team at OpenAI employs a radically different approach. They begin by casting a wide net: “let's list out all of the things that we think we are interested in doing for the next year or two.” Then comes the critical step: “Let's prototype all of them, decide which things are ready now, and then just let the others sit and bake.”

This isn't about discarding ideas, but about patience and constant re-evaluation. The baked ideas aren't forgotten. Instead, “every time there's like a new leap in models, let's try that thing again with it swapped out.” This iterative cycle acknowledges the unpredictable, exponential growth of AI capabilities as the primary driver of product viability.

The real-world impact of this strategy is clear with the OpenAI Codex app. Ambrosino is “very confident that the Codex app that we released in February, if that had been ready in November, it would have absolutely failed in the market and that that the only difference was the models between November and February.” The core lesson: sometimes, the model isn't capable yet, not that your idea is inherently flawed. He advises against being “too AGI-pilled for the moment,” meaning don't prematurely bet on future model capabilities, but keep those ambitious ideas in the oven.

Where This Breaks Down

This "prototype and bake" method isn't for every team or every product. For smaller, bootstrapped startups, the overhead of prototyping dozens of features that might not ship for a year or two can be a significant drain on limited resources. It requires a certain level of engineering capacity and belief in the long game.

The model also assumes a continuous, albeit unpredictable, stream of "model leaps." If AI progress plateaus for an extended period, teams could find themselves with a large backlog of "baked" features that never become viable, leading to wasted effort and potential demotivation. It also requires a high degree of judgment to distinguish between a feature that's genuinely awaiting better models and one that's simply not a good idea, regardless of AI capability.

Finally, managing customer expectations becomes a challenge. How do you articulate your product vision and future while knowing that many desired features are dependent on external AI model advancements outside your direct control? This requires transparency and a customer base willing to embrace a more fluid roadmap.

What to Do With This

For your next AI product roadmap session, ditch the detailed quarterly feature commitments. Instead, dedicate an hour to brainstorming a list of 10-15 ambitious, even "AGI-pilled," AI features you dream of building in the next 12-18 months. Over the next week, have your team build the simplest possible prototype for each. At your follow-up meeting, ruthlessly identify the 2-3 features that genuinely provide value today with current models, and immediately put the rest into a backlog labeled "Bake for Model Leap."