GPT-5.4 Solves 60-Year-Old Math Problem as OpenAI Abandons Manipulated SWE Bench
AI Research & Scientific Breakthroughs
GPT-5.4 Solves 60-Year-Old Math Problem: An amateur mathematician successfully solved Erdős Problem #1196, a challenge that remained unsolved for six decades, using GPT-5.4. The AI proposed a novel mathematical approach that led to the breakthrough, highlighting the potential of large language models to assist in high-level scientific and mathematical discovery.
AI Infrastructure & Hardware Optimization
AMD Hipfire Inference Engine Released: AMD Hipfire is a new inference engine specifically optimized for AMD GPUs, utilizing a unique MQ4 quantization method to achieve significant speed improvements. The project includes a repository of optimized models on Hugging Face, expanding the ecosystem for high-performance local AI inference on non-NVIDIA hardware.
AI Benchmarking & Evaluation
OpenAI Retires SWE Bench Due to Benchmark Manipulation: OpenAI has announced it will no longer use the SWE Bench benchmark for model evaluation, citing concerns that the metric has become "benchmaxxed" or manipulated. This decision reflects growing industry concerns regarding models being overfitted to specific benchmarks rather than developing genuine general capabilities.