When you drop a top-tier AI—famously good at predicting things—into a fully realistic, chronologically replayed "news documentary," does it behave like a prophet or like a flustered intern?
In practice, our proudest AIs behave closer to the latter. 📉
For a long time, we judged AI intelligence with "static knowledge" questions: "What is the capital of France?" or "Write a sorting algorithm." That knowledge is frozen and already baked into training data—like testing a student who has memorized an entire history textbook.
But in May 2026, a major arXiv paper ("FutureSim: Replaying World Events to Evaluate Adaptive Agents") revealed an uncomfortable truth: AI suffers a cliff-edge drop in capability when handling the "unfolding future." 🎢
Researchers built a cyber version of the Truman Show for AI, code-named FutureSim.
What Is FutureSim? 🎞️
Feynman once said: "If you can't predict how a system evolves, you haven't truly grasped its laws."
FutureSim doesn't quiz AI on the past. It performs a kind of "time travel":
1. Go back in time: It selects three months of real global news flow, from January to March 2026. 2. Set the starting point: It places the AI at January 1 and tells it: "Your knowledge of the world ends yesterday. From now on, you must handle 'future challenges.'" 3. Precise replay: The system feeds real, chronological news items one by one. The AI must use this continuously updating information to forecast next week's international situation, market swings, or tech breakthroughs.
Three Painful Moments in This "Reality Show" 💔
Using Feynman-style intuition, here are the fatal weaknesses the test exposed:
1. It's a "dead brain," not "living intelligence" 🧠
Even the strongest model tested (e.g., GPT-5.5) reached only 25% accuracy in this future-simulation contest. AI is good at summarizing "what has already happened," but extremely poor at dynamically revising its "world model" with new evidence.2. The more news it reads, the more confused it gets? 🌀
This is the eeriest finding. Mathematical analysis showed that many AIs' Brier scores (a forecasting calibration metric) were actually worse than blind guessing. Meaning: the more news consumed, the more overconfident the AI became. It grabs one-sided new clues and charges down the wrong path to the end.3. Missing long-distance endurance 🏃♂️
Processing three months of news requires remembering thousands of details and making thousands of tool calls (to check historical context). Current AIs are like runners without stamina: mid-race, they start rambling as the "context gets too long" or the "chain of logic breaks."Why This Paper Matters 🚀
It marks AI evaluation's transition from a "static museum" to a "dynamic battlefield."
Feynman championed "truth through practice." This paper tells us: a true AGI should not be a library that recites books, but an observer that beats in rhythm with the world's pulse.
If we want AI to help predict climate change or prevent financial crises, it must first learn to stay calm in a tsunami of information—and to overturn yesterday's conclusions whenever new evidence demands it.
Summary
Knowledge is past-tense; intelligence is future-tense. ⏳
FutureSim throws cold water on the fantasy that "big models can do everything." It reminds us: measured against the scale of time, AI is still a child.
Next time an AI claims to be omniscient, ask it a FutureSim-style question: "If you don't know what happens tomorrow, how can you be so sure about the day after?"
Truth is not dogma carved in stone, but the courage to keep self-correcting in an unpredictable flow. That is the ultimate lesson in "dynamic intelligence" from 2026's time-travel experiment. 🎓✨