Why AI Products Fail: 9 Hard Truths for Builders

TL;DR

Chat models are a weakly intelligent layer, good enough to start a task but too isolated from your day-to-day data to finish serious work. Real value comes from multi-turn conversations and heavy upfront planning, which delivers power-law returns because AI magnifies good plans and worsens bad ones.

Transcript

Building with AI is hard. Every time we prompt, we're building something. Even if it's as simple as a conversation with AI. These are the nine most overlooked lessons that I have seen come out again and again and again as I have coached people through the journey of getting from the chat GPT interface to actually vibe coding and building apps with ... Read More

Key Insights

• Chat is a weakly intelligent layer because its true intelligence depends on data inputs, and most chat models are strikingly isolated from the day-to-day data environment users operate in. It is good enough to start a task but not to finish serious, precision-heavy work.
• The weakly intelligent layer eats casual-use tool builders alive because people are habitually addicted to AI that is just good enough for most things. This incentivizes serious, precision-requiring tools and disincentivizes casual-category tools.
• Real intellectual work happens in multi-turn conversations, not one-turn answers. A good prompt acts as an anchor at the top that shapes the thread, while the back-and-forth refinement is where the needed intelligence gets surfaced.
• The building vocabulary is the bottleneck, not the AI. People get stopped by 2000s and 2010s-era building systems such as databases, secure data transmission, business logic, transactions, login, and custom integrations rather than by the AI itself.
• Planning delivers power-law returns because AI magnifies good plans and catastrophizes bad ones. Doubling planning time, such as 20 minutes instead of 10, yields far more value since you have more intellectual horsepower behind execution.
• There are three classes of building tools: Cursor, a dedicated AI-powered development environment; Claude Code, a terminal you invoke Claude on to build in the background; and prompting build tools like Lovable that abstract away underlying complexity.
• Conversations should become a fundamental unit of computing, potentially taking the place of the file. Files may remain as an underlying substrate, but systems need to let you build over the top and hide the underlying complexity.
• ChatGPT may be the Instagram of this era, showing strong consumer stickiness off a new underlying technology just as Instagram lived off the mobile phone. This suggests weakly intelligent AI is surprisingly sticky for casual application.

Questions & Answers

Q: What does it mean that chat is a weakly intelligent layer?

Chat is weakly intelligent because the true intelligence of the system depends on the data inputs, and most chat models are strikingly isolated from the data environment you operate in day-to-day. This makes chat good enough to be a task starter, letting you get going quickly, but not good enough to finish the job. Anyone working seriously with AI does not finish work in the chat interface; they may start there but move elsewhere to get the job done.

Q: Why should non-vibe-coders care about these lessons?

You care because this shapes the entire strategic landscape and also shapes the way you engage with chat today, even if you are just a casual chat interface user. Understanding that chat is a weakly intelligent layer, that value lives in multi-turn conversations, and that planning delivers power-law returns changes how you use AI tools regardless of whether you build apps. These lessons come from coaching dozens and dozens of people from the ChatGPT interface to actually building apps with AI.

Q: Why is multi-turn conversation more valuable than one-turn?

AI is reinforcement-learned and architected for one-turn conversations, so thinking in single turns feels natural, but the value is in multi-turn. A good prompt is not designed to get a one-shot answer for the whole thing. Instead, people using AI seriously treat the first message as an anchor prompt that shapes the parameters of the thread, and the thread itself exposes the intelligence needed to do interesting work through the back-and-forth refinement. That conversation is where you find and refine real intellectual work.

Q: What actually stops most people from vibe coding?

Most people who start to build do not get stopped by AI; they get stopped by 2000s and 2010s and earlier-era building systems. The fundamental way we code has not evolved much: building still revolves around having a database, transmitting data securely to and from it, and doing operations with business logic against it. When people run into trouble, it is almost always that they are struggling to build a data-driven application, whether the transactions piece, login, or some custom integration they need.

Q: How much does planning matter when working with AI?

Planning provides enormous leverage, and people dramatically underestimate AI precisely because they do not understand it. If you plan a conversation for twice as long, for example 20 minutes instead of 10, you get far more value because you took time to think about the conversation you wanted to have. With AI, power-law returns are accelerated, so things that were true before are even more true now. Get the planning right or you end up in really wrong territory.

Q: What are the three classes of AI building tools?

There are three basic classes. First is Cursor, a dedicated development environment powered by AI. Second is Claude Code, a terminal you can invoke Claude on, where Claude builds for you in the background and you do not really touch the files. Third are AI-powered prompting build tools like Lovable. Lovable is aggressively scaling up its capability set, enabling you to write more and more complex applications and finish them with the power of underlying models scaling up.

Q: Why might conversations replace files as a unit of computing?

If we want to enable the power of AI agents to build across conversations, we have to make conversations a fundamental unit of computing, which has not been done yet. The conversation is due to take the place of the file. There may still be an underlying substrate of files because that is a surprisingly sticky workspace, but systems need to be enabled that let you build over the top and hide the underlying complexity, like opening a refrigerator door to find food ready and starting to cook.

Q: Why is ChatGPT compared to Instagram?

Just as Instagram lived off the mobile phone and became the runaway success when the iPhone took off, ChatGPT may be the Instagram of this era, a wide consumer success story living off the top of a new underlying technology. ChatGPT is showing that kind of consumer stickiness with its 800 million users, suggesting that weakly intelligent AI is surprisingly sticky for casual application. This matters because it also suggests upside for people building serious tools on top of that layer.

Summary & Key Takeaways

• Chat is not a specialized tool but a weakly intelligent layer whose intelligence depends on data inputs that most chat models are isolated from. It is good enough to start tasks but not finish serious precision work, so real craftspeople move elsewhere to complete the job.

• Most people think in one-turn conversations because AI is reinforcement-learned for that, but the value lies in multi-turn threads. An anchor prompt shapes the parameters, and the back-and-forth refinement surfaces the intelligence needed for real intellectual work like complex documents or code.

• Planning provides outsized leverage because AI accelerates power-law returns, magnifying good plans and catastrophizing bad ones. Building tools split into three classes, Cursor, Claude Code, and Lovable, while conversations are due to become a fundamental unit of computing, potentially replacing the file.