Today it's 1M tokens, tomorrow it's 2M. LLM vendor keynotes increasingly sound like expensive psychological therapy—as if stuffing an entire library into the chat box means the AI can instantly earn you a four-year law degree.
But I suggest you stop that fantasy right now. If you've actually tried feeding a 10,000-word business document to Deepseek or Claude, what you usually get isn't deep insight—it's hallucinations, forgetting, logic gaps, and incoherent nonsense.
This isn't because the AI isn't smart enough. It's because you insist on feeding grass to a Ferrari. 🐄🌱
Attention Is a Brutal Zero-Sum Game
In the Transformer architecture, total attention allocation is fixed at 100%. When you input only 100 characters, each keyword can get 20% of the weight. But when you stuff in 10,000 words, most constraints get diluted down to 0.001%.
And it gets worse: current models have a natural U-shaped memory preference (Lost in the Middle). Put your most critical constraint at word 4,000? There's up to an 80% chance the AI completely ignores it. At that point, the AI is like a brain getting drunk at a party of ten thousand people—it only clearly hears the persona setup at the door and the formatting requirement you shout on the way out. 🥴
The Cargo-Cult Prompt
Many people write long prompts like coconut-shell headphones. After WWII, South Pacific islanders wore coconut-shell headphones beside runways, waiting for planes to deliver cargo—that's cargo cult. You write 10,000 words that look logically rigorous and even sincerely pleading, but 95% of it is low-entropy filler: "I think," "hopefully," "probably."
To a large language model, this isn't feeding material—it's poisoning it. A Ferrari needs refined, structured gasoline; your prompt is just weeds dumped in the fuel tank.
Here's the uncomfortable truth: the more sentimental and verbose your writing, the dumber the AI gets. 🧠📉
Compile, Don't Chat
If you want a beast like Deepseek V4 to truly work for you, you must abandon "chat thinking" and adopt systems-engineering thinking.
You have to "compile" your prompts the way you write code. Use XML tags to create physical boundaries, cleanly separating read-only <Context> from mandatory <Workflow>. If you don't have the patience to build this kind of "spatial lockdown," your answers will forever be generic boilerplate that a search engine could find.
Input ≠ Output
More brutal still: an input of 1 million tokens doesn't mean an output of 1 million.
Though theoretically capable of producing hundreds of thousands of words, under the curse of probability, AI logical coherence typically breaks around 8,000 tokens. Every extra character you ask for exponentially increases the risk of it talking nonsense. 📈
My Bet
If by 2026 you're still using rambling natural-language prose to steer AI, I guarantee you'll get nothing but wasted electricity and tokens. Meanwhile, the people who learned to "compile instructions" will have replaced your entire legal department with AI.
If you disagree, keep writing your 10,000-word essays. See you in 2027. 🤝
---
References
- Technical analysis: *"Lost in the Middle: How Language Models Use Long Contexts"*, Stanford/UC Berkeley.
- Architecture research: *"The Transformer Attention Bottleneck: A Zero-Sum Game for Contextual Information"*, 2025.
- Engineering guide: 莫里AI, *"为什么 Deepseek 100万上下文写两万字都会胡说八道？"*, 2026.
- Logic standards: *"XML Tagging as a Physical Boundary for Agentic Reasoning"*, Deepseek Research, 2025.