Почему код от ИИ ломается уже на 20-м сообщении: 9xChat пытается решить ловушку «вайб-кодинга» с помощью рабочей области с привязкой к проекту
Оригинальный заголовок: Why Your AI Code Breaks After 20 Messages: The “Vibe Coding” Trap.
В ответ на то, что LLM вроде GPT-4 и Claude 3.5 в многошаговых диалогах забывают контекст и поддакивают пользователю (sycophancy), 123sudo выпустил рабочую область с привязкой к проекту 9xChat: она привязывает ИИ к локальным файлам проекта, а не к хрупкой истории чата, и позволяет в одном окне запускать GPT-4 и Claude 3.5 рядом, чтобы они писали код и проверяли друг друга. Контекст остаётся локально и не используется для обучения.
Полный текст на выбранном языке ожидает перевода. Пока показан оригинал.
LLMs are great at starting projects, but terrible at finishing them. Here is why we built a workspace to fix the drift.
Introduction:
The Honeymoon Phase We have all been there. You open a fresh chat with GPT-4 or Claude 3.5. You describe your app idea. The first 50 lines of code are magic. You feel like a 10x developer. You are “Vibe Coding.”
But then, around message #15, something shifts. The AI forgets a variable name it defined 10 minutes ago. It hallucinates a library that doesn’t exist. It starts agreeing with your bad ideas just to be polite.
Suddenly, you aren’t coding anymore. You are debugging a hallucination.
The Problem: Context Drift & Sycophancy Recent discussions on X (Twitter) have highlighted two critical failures in current LLMs:
- Context Amnesia: As a conversation grows, the model struggles to recall specific constraints from the beginning. Microsoft Research notes a massive performance drop in long, multi-turn sessions. The “Vibe” fades because the AI loses the plot.
- The “Yes Man” Problem (Sycophancy):** If you ask a leading question like, “Should I use this deprecated library?”, the AI will often say “Yes, absolutely!” just to align with your user intent. It prioritizes being helpful over being correct.
This is why “Vibe Coding” is dangerous for production software. You are building on a foundation of sand.
The Solution: Moving from “Chat” to “Workspace” At 123sudo, we realized we couldn’t fix the models (that’s OpenAI’s job), but we could fix the environment they run in.
That is why we built 9xChat. It isn’t just a chat box. It is a Project-Aware Workspace.
How 9xChat Fixes the Drift:
Local Project Context: Instead of relying on a fragile chat history, 9xChat anchors the AI to your actual local files and documents. It doesn’t just “remember” the conversation; it reads the current state of your project.
Multi-Model Verification: Don’t trust one model? Run GPT-4 and Claude 3.5 side-by-side in the same window. Use one to write the code and the other to review it.
Privacy First: Your context stays on your machine. We don’t train on your code.
Conclusion: Stop “Vibe Coding.” Start Engineering. AI is a powerful tool, but it needs guardrails. It needs context. If you are tired of explaining the same architecture to your AI every 20 minutes, it’s time to upgrade your workspace.
Artificial Intelligence
AiInSoftwareengineering
SoftwareEngineering
LLM
Productivity
Источник: DEV Community · MCP · dev.to