AI is Breaking: Why Ilya Sutskever Just Declared the End of the Scaling Era

The AI Hype vs. Reality

• 💡 Initial widespread belief suggested AI would solve major global problems like climate change and world hunger.
• 🚀 Many experts predicted AI would lead to a utopian future and serve as a co-pilot for humanity.

The Gauntlet: Testing AI Limits

• 🎯 Researchers created "the gauntlet," a series of three brutal exams designed to find the breaking point of advanced AIs.
• 🧪 These tests included Agent Bench for basic digital tasks, Scientific Discovery Evaluation for scientific reasoning, and Humanity's Last Exam as the ultimate challenge.

Shocking AI Performance Failures

• 🧼 GPT-4, the supposed "heavyweight champion," failed a simple intern test to find a soap bar, demonstrating a lack of basic task execution despite complex planning.
• 📉 Top models like Google's Gemini 3 Pro and GPT 5 Pro scored dismally low (37.5% and 31.6%) on "Humanity's Last Exam," indicating a significant failure.
• 🎓 AI's performance on the science test was worse than an ordinary undergraduate student, highlighting its inability to perform scientific tasks effectively.

Core Reasons for AI Breakdown

• ⚠️ Researchers identified four main failure categories: invalid format, invalid action, task limit exceeded (getting stuck in loops), and hallucinations (making things up).
• 📊 OpenAI's internal research revealed their best model has only a 42.7% success rate, meaning it is wrong more often than it is right.

The Dangerous Implications of Flawed AI

• 🚨 AI is described as a powerful, world-changing tool but also a "confident and convincing liar" that is unaware of its own falsehoods.
• 🏥 This flawed technology is already deployed in critical sectors like hospitals (inventing medications) and police departments.
• 🧠 Users are urged to be smart enough to catch AI's lies and recognize its limitations, remembering the "soap" and "failing grades."