OpenAI adds new roles at Dublin EU headquarters

A fork of ExLlamaV3 adds native training in plain PyTorch without transformers, letting users fine-tune quantized models across 2-8 bits per weight. The author says it has been in development for several months and is now stable enough to share.

ReplaySSM stores SSM state as a checkpoint plus a buffer of recent inputs, separating output computation from state updates. NVIDIA says SSM layers can bottleneck latency up to 100K tokens of context at serving batch sizes. It is integrated into TensorRT-LLM, vLLM, and FlashInfer, and runs in production in GLM-5.3-Flash and Kimi-K3.

Hrishi Olickel of Southbridge AI recounts a pipeline that stripped minus signs from loan amounts, turning a net $784,000 into $5.3 million of debt, with multiple models agreeing on the wrong result. He argues data transformation needs more than a universal schema or a confident agent.