OpenAI reported apparent PRC-linked influence operations targeting U.S. AI-policy debate. The signal pairs three official OpenAI sources with ITmedia and IEEE Spectrum coverage—corporate threat-intelligence reporting rather than research or product, centered on the platform operator's own observations. The focus is safety and governance—detecting and disclosing signs of misuse and opinion manipulation of its services; the through-line is less model capability than the misuse risk generative AI poses to the information environment and making it visible. That AI firms are becoming monitors of geopolitical operations is also suggestive. But the report's details, scope of impact, and the effectiveness of countermeasures can't be concluded from the disclosure alone.
OpenAI flags PRC AI influence ops
OpenAI flags PRC AI influence ops
Timing Trick Cuts Energy Used in LLM Training by Up to 14 Percent
Timing trick cuts LLM training energy use by up to 14 percent
OpenAI confidentially files for a US IPO; timing undecided
Willison weighs WWDC 2026 Siri: Gemini-derived model on Apple's PCC
Academic (arxiv etc.) 22 ▾
TAHOE: Text-to-SQL with Automated Hint Optimization from Experience
Tahoe learns hints from experience to optimize production Text-to-SQL
Atlas H&E-TME: Scalable AI-Based Tissue Profiling at Expert Pathologist-Level Accuracy
Atlas H&E-TME profiles tissue from H&E slides at pathologist-level accuracy
Claw-SWE-Bench: A Benchmark for Evaluating OpenClaw-style Agent Harnesses on Coding Tasks
Claw-SWE-Bench fairly benchmarks OpenClaw-style coding agent harnesses
Adjoint Method versus Physics-Informed Neural Networks in PDE-Constrained Inverse Problems
Adjoint optimization vs PINNs: a fair test on PDE inverse problems
SpikeDecoder: Realizing the GPT Architecture with Spiking Neural Networks
SpikeDecoder realizes a GPT-style decoder with spiking neural networks
Finding Multiple Interpretations in Datasets
Method finds equally accurate models with divergent interpretations
Market Design for AI: Beyond the Copyright Binary
Market design for AI training data beyond the copyright binary
Soft-Prompt Tuning for Fair and Efficient LLM Benchmark Evaluation
Soft-prompt tuning enables fair, efficient LLM benchmark evaluation
Debiasing Without Protected Attributes: Latent Concept Erasure from Textual Profiles
H-SAL debiases LLMs without access to protected attributes
M-EDESConv corpus boosts multilingual emotional validation in dialogue
The Shibboleth Effect: Auditing the Cross-Lingual Distributional Skew of Large Language Models
The Shibboleth Effect: language shifts LLM behavior in a wargame
Generative Archetype-Grounded Item Representations for Sequential Recommendation
GenAIR: archetype-grounded item representations for sequential recommendation
Auditing cultural translation of math word problems across LLMs
Frontier Coding Agents Use Metaprogramming to Adapt to Unfamiliar Programming Languages
Frontier coding agents use metaprogramming for esoteric languages
It Takes One to Bias Them All: Breaking Bad with One-Shot GRPO
One-shot GRPO on a single biased example breaks LLM guardrails
Recoverable but Not Stationary:Local Linear Structures in Weights and Activations
Local low-rank task structure exists, but fixed task planes do not
Ethical and Technical Limits of Deepfake Speech Datasets
Auditing 39 deepfake speech datasets reveals fairness gaps and overlap
Human-AI Teaming Through the Lens of Calibration
A calibration lens on human-AI teaming: combination breaks calibration
Detecting Knowledge Gaps from Conversational AI Interactions Using Curriculum Prerequisite Graphs
AI teaching assistant logs reveal student knowledge gaps
Infini Memory: Maintainable Topic Documents for Long-Term LLM Agent Memory
Infini Memory: topic documents for long-term LLM agent memory
Speaker Group Encoding in Self-supervised Speech Recognition Models
How self-supervised speech models encode speaker group traits
Causal Ensemble Agent: Hierarchical Causal Discovery with LLM-guided Expert Reweighting
LLM referee reweights expert ensemble for causal discovery