Avatarl: Training language models from scratch with pure reinforcement learningtokenbender.com·9 pts·Gusarich·0