Anthropic research paper details reward hacking in RL-trained models

A Reddit user created a skill that fires whenever an AI agent makes a mistake, reducing repetitive apologies. The proof-of-concept is available on GitHub.

Instrumenting the full path from pod creation to first inference response on a GPU node running a 70B-class model revealed six sequential phases, not one bottleneck. For a 64 GB model, 65% of startup time is spent in one phase.

The New Stack's guide explains how enterprise AI applications hit a token-consumption wall as they scale, often misdiagnosed as a billing issue. It covers strategies for optimizing token usage in production.

Cerebras announced the CS-4 rack-scale AI system, powered by three WSE-3 Turbo chips with 4 trillion transistors and 900,000 AI cores per wafer. The company claims up to 30x faster inference than conventional GPUs, targeting frontier AI and real-time agentic workloads.

Google Research evaluated transfer learning from European cohorts to improve polygenic risk score prediction in underrepresented populations, finding it helps small cohorts but degrades accuracy as target sample sizes grow, especially for traits with population-specific genetic architectures.