A new autonomous programming agent developed by @shawnup, utilizing OpenAI's o1 model, has achieved a state-of-the-art score of 64.6% on SWE-Bench Verified. This accomplishment, which marks a significant advancement in the field, was made possible through the use of the Weights & Biases Programmer and the Weave toolkit, alongside continuous iterations. The results have garnered positive recognition within the AI community, with various users congratulating @shawnup for this milestone. Additionally, DeepSeek-V3, an open large language model (LLM) from DeepLearningAI, has surpassed notable benchmarks, including Llama 3.1 405B and GPT-4o, particularly excelling in coding and math tasks. This model employs a mixture-of-experts architecture with 671 billion parameters, activating only 37 billion at any given time, and has been trained at a low cost, indicating advancements in AI capabilities from China.