Earendil released experimental Pi Durable alongside Pi 1.0: an open TypeScript harness for long-lived agents with checkpointed task recovery, concurrent conversations, steering from multiple clients, and pluggable storage/execution environments. Novel because it focuses on durable/multi-user agent infrastructure rather than another coding-agent UX; the launch drew 359 HN points and 45 comments, though the package is explicitly experimental and its API may change. Sources: https://earendil.com/posts/pi-durable/ ; https://news.ycombinator.com/item?id=49925969
ServiceNow CoreAI published AutoSynthData, a method for generating executable, verified enterprise-agent training tasks from a target model’s capability gaps, then iteratively refining the curriculum. In EnterpriseOps Gym experiments, ServiceNow reports +7.2 Pass@1 points on Hybrid and +8.41 points on ITSM after synthetic-data fine-tuning; these are vendor-authored benchmark results in the tested environment, not independent replication. Sources: https://huggingface.co/blog/ServiceNow-AI/autosynthdata ; https://huggingface.co/ServiceNow-AI/EnterpriseOps-Gym
Ai2 released Olmo-core 3, open mixture-of-experts training infrastructure intended to scale to trillion-parameter models, with code and a technical report. Ai2 reports about 2.7× throughput versus its previous implementation on an 8-GPU B300 benchmark and a 1.2T-parameter system configuration across 512 GPUs; the latter uses random routing and demonstrates systems scale, not a trained model’s capability. Relevant to the model radar as an open training-stack release preceding Ai2’s next MoE model. Sources: https://allenai.org/blog/olmocore3 ; https://github.com/allenai/olmo-core ; https://huggingface.co/blog/allenai/olmocore3