Is it true that the latest AI models score poorly on physics exams? Recent research suggests AI's true capability might be far higher than we previously thought.
We explore why inspection tools (evals) that verify AI functionality sometimes send false warnings and why evaluating AI is so challenging.
We explore why AI judges, which evaluate the accuracy of AI-generated medical records, struggle to identify missing information, along with the reasons and limitations behind this issue.
Why do AI models that speak like humans give strange answers to math or logic problems? We examine the surprising limitations of Large Language Models (LLMs) and the reasons behind them.
It has been revealed that a significant number of GPU kernel codes written by AI contain defects. We introduce a new 'contract-grade' verification tool to solve this problem.
We analyze the reality and actual efficiency of RTK technology, which is touted to dramatically reduce token costs incurred when using AI coding tools.
We explain why China's high-performance AI 'Kimi K3' is frequently compared to Anthropic's Claude and the secrets behind their surprising similarity.
A recent AI competition has sparked controversy after entries dubbed 'AI Slop' (low-quality data) took home a large cash prize. Did the AI truly understand the problem, or was it just a fluke?
Explore the potential for advanced AI to be exploited in cyberattacks and the new security evaluation frameworks designed to prevent them.