According to AI at Meta on X, Meta trained its Muse models, including Muse Wonder, on the Prometheus supercluster, reporting 85%+ pretraining goodput, 90%+ RL goodput, and about 47K GPU hours mean time between interruptions.

According to gdb, OpenAI’s first NVIDIA Vera Rubin racks are live, scaling compute for next frontier pretraining, per Uday Ruddarraju on X.

According to StanfordAI Lab, residual repetition after deduplication can waste up to 33% of FLOPs, with worst-case patterns predictable by model size.

According to Ethan Mollick on X, researchers released an LLM trained entirely from scratch on over 28,000 Victorian-era British texts (1837–1899) sourced from the British Library dataset, positioning it as fundamentally different from generic models merely roleplaying a Victorian persona. As reported by Ethan Mollick, the model’s domain-native pretraining enables authentic period syntax, vocabulary, and cultural references, which can improve historical dialogue agents, archival assistants, and stylistically faithful content generation. According to the British Library dataset description cited by Ethan Mollick, the corpus scale supports robust language modeling for 19th-century English varieties, suggesting opportunities for museums, publishers, and edtech to build specialized chatbots, curriculum tools, and literary restoration pipelines. As noted by Ethan Mollick, training from scratch versus fine-tuning reduces modern-language interference, potentially yielding better retrieval-augmented generation for heritage collections and more accurate period entity disambiguation.