Sep 7, 2025

• GPT-5, dubbed the "Research Goblin," excels at search tasks, outperforming traditional search engines by efficiently synthesizing information from the web. This capability allows users to delegate complex queries and receive comprehensive answers, enhancing both casual curiosity and serious research efforts.

• Performance improvement: The new kernel achieves 288.3 TFLOPs, marking a 58x enhancement over the initial implementation, yet it still operates at only 16.4% of cuBLAS performance due to global memory access limitations.

• Knowledge Distillation (KD) enables GPT-2 to effectively generate SQL queries from natural language by learning from the larger Qwen2-7B model, demonstrating that smaller models can be trained for specific tasks without the need for massive resources.

• Baseten's $150M Series D emphasizes their commitment to enhancing inference infrastructure, targeting low latency and throughput optimization, with a valuation of $2.1B.

• BitTransformerLM is a 16M parameter model that processes raw binary data directly, bypassing tokenization, which could revolutionize how transformers are built and trained.