All news
6 articles- paperEmergent Mind·25 Oct 20264.2Finds numerical bugs in standard softmax and cross-entropy implementations, like softmax collapse, that change training dynamics, and proposes fixes.
Also in arXiv (systems, data, ML)
- newsInfoQ·26 Oct 20263.9Google and Meta open-sourced
TorchTPU, a native PyTorch backend for TPUs, and Meta plans to move larger LLM workloads onto TPUs; AWS releasedTorchNeuronfor Trainium the same week. - blogSebastian Raschka·24 Oct 20263.8A full video course on building reasoning LLMs from scratch, covering inference-time scaling and RL training.
- paperHugging Face Daily Papers·20 Oct 20263.6How RL papers get into seven open-source LLM training frameworks.
- newsTechCrunch·24 Oct 20263.3OpenAI will train models on cached and simulated copies of the web instead of the live internet, citing risks from capable models.
- paperIEEE Xplore·1 Nov 20263.1Survey of low-precision training for LLMs, including FP8 and lower formats.