This is today's edition of The Download, our weekday newsletter that delivers a daily digest of technology's most pressing developments.
AI Models Stumble on These Intelligence Tests—Can You Do Better?
Puzzles and games have always been central to AI development. The term “machine learning” gained traction in a 1959 article describing an algorithm that learned to play checkers. Chess and Go further cemented games as key benchmarks for AI capabilities.
On raw puzzling prowess alone, AI is advancing rapidly. In late 2024, even the most sophisticated models could solve only 18% of the notoriously challenging New York Times Connections puzzles. By early 2025, some models approached near-perfect scores. Yet by 2026, the frontier has shifted: today's models excel at pattern recognition but still struggle with tasks requiring abstract reasoning and real-world common sense—areas where human intuition often prevails.
Puzzles, however, do more than chart AI's progress. Observing where models excel, stumble, and fall short of human performance offers a window into the technology's inherent strengths and limitations. For instance, models may ace a logic grid but falter on a riddle embedded in cultural context, revealing gaps in their training data and reasoning frameworks.
We've curated seven puzzles that have famously confounded AI systems at one stage or another. Now, it's your turn to test your own wits. Take our quiz to see if you can outsmart artificial intelligence.
—Grace Huckins
Also in today's news: OpenAI has restricted access to its newest model after internal evaluations flagged it as a “critical” cyber risk, a decision that underscores the growing tension between AI capability and safety. In other developments, researchers have proposed a novel propulsion pathway toward Alpha Centauri, our nearest star system, potentially cutting travel time from millennia to decades—a bold vision that could redefine interstellar exploration.