The ladder broke all afternoon, feeling like I've lost connection with the world.

The most advanced batch of AI models writes code that isn't bad, with module-level performance far exceeding the average level of human programmers. If the code produced is poor, first check which model was selected, whether the context is sufficient, and if the prompts need optimization. Cross-module code, limited by the context window length, may require human assistance for design and planning. If the project structure is reasonable, AI can also reuse existing code to maintain DRY.

After careful consideration, the reason AI writes poor code is not that it can't write, but that it can't read. Humans look at many seemingly unrelated things to implement a feature, but it's precisely these seemingly unrelated things that provide the possibility of keeping the project DRY. For AI to become a pillar, it must first learn this.

Don't even mention the Bodhisattva. If you were out in the wild and about to starve... and you saw offerings at a grave by the roadside. You bow your head and say, "Sorry, I'm about to starve, can I take an offering to eat?" I don't think anyone or any ghost would make things difficult for you...

Why does it feel like AI engineering research is becoming more focused on language philosophy?

I deeply agree: 1. Multi-agent parallel collaboration is less stable than single-threaded; 2. RAG is unreliable and worse than traditional retrieval; 3. The more instructions in the prompt, the less the model knows how to choose. —— The original translation is as follows —— On the road to building AI agents, our team @Cline has identified three "thought viruses." These so-called "thought viruses" are enticing ideas that sound brilliant but are completely impractical in practice. The three viruses are: * Multi-Agent Orchestration * Retrieval Augmented Generation (RAG) * More instructions = better results Let's explore them! 1. Multi-Agent Orchestration The kind of scene from a sci-fi movie—"back-end agents, supply agents, analysis agents, command agents" dispatching a large group of sub-agents, and then summarizing the results—sounds really cool. But the reality is that most useful agent work is single-threaded. Complex collaboration processes rarely bring real value and often create chaos instead. It's important to know that just getting the model to work stably in a single thread is already quite difficult, let alone handling those parallel collaboration logics. This not only increases implementation complexity but also makes the model's understanding and decision-making process exceptionally complicated. 2. Building agents with RAG RAG, or Retrieval Augmented Generation, is also a thought virus. It looks powerful in theory, but in practice, especially in agent scenarios, sometimes even a basic text search command like GREP is more useful than it. Why does the halo of RAG fade in actual agent workflows? Because the information retrieved is often fragmented, preventing the model from forming a coherent and useful "understanding." A better approach is almost always: let the model list the files itself, search in a grep-like manner, and then open and read the entire file (just like humans do). The @Cline team started doing this early on, and later we saw @Amp — Research Preview and @Cursor also shift to this more pragmatic approach. 3. More instructions = better results There is a widespread misconception: piling more "instructions" into the system prompt will make the model smarter. This is completely wrong. "Watering down" the prompt will only confuse the model, as more instructions often lead to conflicting suggestions and information overload. The result is that you have to keep patching the model's various strange behaviors like playing a "whack-a-mole" game, rather than getting truly useful outputs. For most cutting-edge models today, the best approach is to not block their path, rather than shouting from the sidelines, trying to steer them in a specific direction. Cherish every word (or Token) you use. In summary, all three of these ideas are highly tempting. If you don't deal with AI all day, you might find them very reasonable—however, that is not the case. Of course, as the capabilities of underlying models improve, our views on these methods may change in the future.