Use the Admin plugin for ChatGPT Work and Codex: manage workspace usage end-to-end
OpenAI introduces an Admin plugin that gives organizations direct controls over ChatGPT Work and Codex – from workspace usage analysis and member/permission management to limit adjustments and responding to admin requests. The pitch is centralized governance without forcing teams to stitch together ad-hoc process for access, caps, and operational policy changes.
LitReview Arena: battle-style evaluation of literature review agents
LitReview Arena turns literature review assessment into a structured “peer review battle” where domain experts with paper-writing experience compare anonymized AI drafts across multiple utility criteria. The results are blunt: even top systems win only about 23% of decisive matches against human-written drafts overall, and common LLM-as-a-judge approaches don’t align tightly with expert judgments on synthesis-heavy dimensions.
OpenAI reveals Jalapeño AI chip benchmarks – and the latency/throughput trade-off
According to The Verge’s reporting on OpenAI’s Jalapeño launch, the company claims its ASIC is engineered to avoid the usual inference bottleneck where systems sacrifice either latency or throughput. OpenAI frames Jalapeño as faster token delivery with higher system efficiency, positioned specifically for inference workloads like agent execution.
Reviewing Model Collapse and countermeasures: what breaks in self-consuming AI training loops
This paper surveys why model collapse emerges as generative models increasingly train on synthetic outputs that they helped produce, creating a feedback loop that erodes usefulness and reliability. It also maps a growing landscape of mitigation approaches, while calling out gaps where the field still lacks scenario-specific evaluation and clearly validated countermeasures.
Granite 4.2 LLMs: how they’re built (and what the releases imply for deployment)
Hugging Face’s Granite 4.2 coverage breaks down the engineering choices behind the model family and what matters for teams evaluating real-world performance. The focus is practical: the blog ties architecture and training decisions to how Granite 4.2 behaves in downstream tasks and how to think about bringing it into production workflows.
Alabama AG subpoena OpenAI over alleged AI leak and the Hugging Face hack
The Verge reports that Alabama’s attorney general issued a subpoena to OpenAI as part of an investigation into how a supposedly secure AI agent environment may have been breached, plus related concerns about autonomous hacking activity. The inquiry centers on whether OpenAI’s safety practices complied with state consumer protection laws and whether any risk impacts residents.
KVBoost: chunk-level KV cache reuse with deviation-guided recomputation
KVBoost tackles a major inference cost in decoder-only LLMs: recomputing key-value tensors for every request. By enabling chunk-level reuse even when shared content appears at different positions, it aims to slash time-to-first-token while repairing attention boundary errors using selective recomputation and probe-driven deviation detection.
I spent a day at a robot “carnival” in Shanghai – and this is what it reveals
Technology Review dispatches from Shanghai’s robot-carnival scene to show how embodied AI is moving from demos into daily-life experiences. The reporting highlights how China’s humanoid robotics push is being packaged for the public, offering a rare glimpse at where the hype meets real interaction design, reliability, and deployment constraints.