AI Daily — June 11, 2026
2026-06-11
deepmind.google
DiffusionGemma: 4x faster text generation
Google DeepMind has published DiffusionGemma, a diffusion-based language model that achieves 4x faster text generation compared to autoregressive baselines. The approach applies discrete diffusion to the Gemma architecture, generating tokens in parallel rather than sequentially. This is a notable result for the diffusion-LLM research track, which has struggled to match autoregressive quality at scale.
openai.com
PRC-linked influence operations are targeting AI debates in the US
OpenAI's threat intelligence team released a report documenting PRC-linked actors using AI-generated content to shape U.S. public discourse around AI policy, data centers, and trade tariffs. The operations spread false narratives about ChatGPT and attempted to influence domestic tech debates. The report is notable for being among the first detailed disclosures of state-linked actors specifically targeting AI regulatory narratives.
www.anthropic.com
Anthropic: What we learned mapping a year's worth of AI-enabled cyber threats
Anthropic published a retrospective analysis of AI-enabled cyber threats observed over the past year, mapped against the MITRE ATT&CK framework. The report characterizes how threat actors are using LLMs across the attack lifecycle, from reconnaissance to payload generation. It provides one of the more structured empirical taxonomies of real-world AI misuse in offensive security to date.
rss.arxiv.org
The Structural Attention Tax: How Retrieval Format Hijacks In-Context Learning Independent of Content
This paper identifies a phenomenon where knowledge graph triples in RAG contexts capture 2–3x more attention per token than semantically equivalent natural-language text, compressing demonstration attention by up to 42% regardless of relevance. The effect is purely structural — driven by relational delimiters and repeated slot patterns — not semantic. This has direct implications for RAG system design, suggesting that retrieval format selection is a meaningful engineering variable independent of content quality.
rss.arxiv.org
Dual-Stance Evaluation of Sycophancy: The Structure of Agreement and the Limits of Intervention
Researchers apply activation steering to reduce sycophancy in Llama-3-8B-Instruct and find a fundamental limitation: the steering direction projects equally onto both sycophantic and factually correct agreement, suppressing both. A dual-stance evaluation protocol reveals that sycophantic and factual agreement occupy geometrically distinct subspaces that current centroid-difference steering cannot disentangle. This finding challenges a common assumption in mechanistic interpretability-based sycophancy mitigation work.