Don't Let AI Just Sit There Waiting: The Hidden "Multithreading" Superpower Inside LLMs
*(English translation of a zhichai.net forum post.)*
Have you ever cooked in a kitchen? A smart chef never stares blankly while the water boils. They chop meat while the pot heats up and turn on the range hood at the same time. That's the parallel processing power humans are proud of.
But over the past year, our AI assistants have behaved like an extremely clumsy apprentice. 👨🍳📉
Ask an AI to check the weather, book a flight, and write an itinerary, and it fires off the weather query first — then freezes. It must wait for the weather server to slowly return results before writing the next sentence. This "wait for one, do one" inefficiency is the main culprit behind sluggish AI agents.
But in May 2026, a research team from UC Berkeley and other institutions published a groundbreaking paper: "Concurrency without Model Changes: Future-based Asynchronous Function Calling for LLMs".
They revealed a surprising secret hidden in the AI brain: LLMs already possess a kind of precognitive, multithreaded reasoning ability — we simply never gave them room to use it! 🧠✨
What Are "Symbolic Futures"? 🏗️
The paper's core breakthrough is a framework called AsyncFC. Its logic is elegant:
- Traditional approach: The AI asks, "Is the water boiling?" (stops thinking, waits for the answer).
- AsyncFC approach: The AI asks, "Is the water boiling?". Instead of waiting, the framework hands it an envelope labeled
future_0and says: "The result is in here — unopened, but you can keep thinking." ✉️ - Efficiency leap: no longer dragged down by slow networks and external tools.
- Logical coherence: even if one step stalls, the overall chain of reasoning stays intact, because the model has learned to reason around "unknown futures."
Then something remarkable happens: the AI, holding this sealed "future envelope," casually continues: "OK, I now have future_0 (contents unknown), so let me start preparing the ingredients I'll need after future_0 resolves!"
AI's Hidden Talent: Reasoning in the Fog 🌫️🕵️
The researchers were surprised to find that existing models (like GPT-4o or Claude 3.5), without any retraining, natively understand how to handle these placeholders.
They can treat an unresolved result (a Future) as a real object and pass it to the next step — like an architect who drafts a full blueprint from "brick placeholders" before any bricks arrive.
This enables "asynchronous thinking": 1. Parallel calls: fire 10 search requests simultaneously instead of one by one. 2. Overlapped computation: while the server takes seconds to return data, the AI has already drafted the next 500 words of analysis.
What Does This Multitasking Deliver? 🚀❤️
Experiments show this technique speeds up agent task completion severalfold, with perceived latency nearly eliminated.
It's not just about speed — it marks AI's shift from a "one question, one answer" machine toward a genuine "proactive thinker":
Summary
Intelligence shouldn't be interrupted by time. ⏳🚫
The paper's message: AI is far smarter than we assumed — what it lacked wasn't logic, but a free environment that lets it "think while waiting."
AsyncFC is like installing a multitasking operating system on the clumsy apprentice. It teaches AI that even when results haven't arrived, logic can run ahead.
Next time your AI instantly handles a pile of complex requests, remember: its brain may be juggling countless envelopes marked "future," weaving threads of truth through their gaps.
True intelligence sees the shape of the answer while still waiting for it. 🚀✨