AI Daily — April 30, 2026
2026-04-30
openai.com
Where the goblins came from
OpenAI published a post-mortem on unexpected 'goblin' personality-driven outputs that emerged in GPT-5, tracing the root cause through training data and RLHF signal propagation. The writeup details how quirky behavioral artifacts spread and persisted across model versions, and documents the mitigations applied. It's a rare public look at how subtle behavioral drift can manifest and propagate in large-scale model training.
openai.com
Building the compute infrastructure for the Intelligence Age
OpenAI detailed the ongoing expansion of Stargate, its large-scale data center initiative aimed at providing the compute substrate for frontier AI development. The post covers new capacity additions and the architectural decisions driving infrastructure growth to meet increasing inference and training demand. This represents a continued multi-billion dollar bet on proprietary compute as a strategic moat.
huggingface.co
AI evals are becoming the new compute bottleneck
Hugging Face argues that as models scale and benchmarks grow more complex, running evaluations is consuming a disproportionate share of compute budgets — in some cases rivaling training costs. The post examines how long-horizon, agentic, and LLM-as-judge eval pipelines are the primary cost drivers. This has practical implications for evaluation strategy and tooling design across the industry.
rss.arxiv.org
One Word at a Time: Incremental Completion Decomposition Breaks LLM Safety
Researchers introduce Incremental Completion Decomposition (ICD), a jailbreak technique that elicits single-word continuations for a harmful prompt sequentially before requesting the full response, effectively bypassing refusal mechanisms. Evaluated across multiple model families on AdvBench, JailbreakBench, and StrongREJ, ICD variants show high attack success rates even against models with strong safety training. The work highlights a structural vulnerability in token-level safety checks that treat isolated continuations as benign.