Today it's a 1M-token context window, tomorrow it's 2M. LLM vendor keynotes increasingly sound like expensive psychological massage — as if stuffing an entire library into the chat box means the AI can instantly finish four years of law school for you.
But stop the fantasy. If you've actually thrown a 10,000-word business document at DeepSeek or Claude, what you typically get back isn't deep insight — it's hallucinations, forgetting, broken logic, and incoherent filler.
This isn't because the AI isn't smart enough. It's because you insist on feeding grass to a Ferrari. 🐄🌱
AI attention is a brutal zero-sum game
In the Transformer architecture, total attention allocation is a hard 100%. When you input only 100 words, each keyword can capture ~20% of the weight; but when you stuff in 10,000 words, most constraints get diluted to 0.001%.
Worse, current models have a natural U-shaped memory bias ("Lost in the Middle"). Put your most critical constraint around the 4,000th word, and there's roughly an 80% chance the AI completely ignores it. At that point the AI is like a drunk brain at a party of ten thousand people: it only hears the persona setup at the door and the format requirement you shout on the way out. 🥴
Cargo-cult prompts
Many people's long prompts are like coconut-shell headphones: post-WWII Pacific islanders wore coconut headphones beside runways, hoping planes would deliver cargo. That's cargo cult thinking. Your 10,000 words may look rigorous and sincere, but 95% of it is low-entropy filler — "I think," "hopefully," "probably."
To an LLM, that's not feeding data — it's poisoning it. A Ferrari needs refined, structured fuel, not weeds dumped in the tank.
The uncomfortable truth: the more sentimental and rambling your writing, the dumber the AI gets. 🧠📉
Use systems engineering, not chat
If you want a beast like DeepSeek V4 to truly work for you, abandon "chat thinking" and adopt systems-engineering thinking.
You must "compile" your prompt like code. Use XML tags as physical boundaries, strictly separating read-only <Context> from mandatory <Workflow>. If you can't be bothered to build this kind of spatial isolation, your answers will forever be generic boilerplate that a search engine could find.
Output doesn't scale with input
Harsher still: 1 million input tokens doesn't mean 1 million output tokens.
Although tens of thousands of words are theoretically possible, under the curse of probability, an AI's logical coherence usually breaks around 8,000 tokens. Every extra character it writes multiplies the risk of nonsense. 📈
The bet
If by 2026 you're still fine-tuning AI with "running commentary" natural language, I guarantee you'll get nothing but wasted electricity and tokens. Meanwhile, those who learned to "compile instructions" will have AI replacing your entire legal department.
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.