The Archive

Google's and Reddit's use of DMCA to fight web scraper is bizarre, expert says.

Telecom expects AI revenue from dark fiber deals and retrofitted data centers.

Anthropic publishes official stance on open-weights model releases, addressing trade-offs between transparency, safety, and competitive positioning.

Microsoft bolstered its AI cybersecurity offerings this week with the launch of its first AI security model and a new security platform.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Reading OpenAI’s account last week of how some of its models broke their containment and hacked into the computer systems of Hugging Face, another AI company, was the first time I got…

ClinFusion: vision-centric MLLM for medical imaging that integrates 2D/3D images with clinical evaluation aligned to radiologist workflows.

TemporalSinkhorn: parallel-in-time algorithm for batching entropic optimal transport computations without speculative output.

Theoretical study of distribution learning from heterogeneous data providers using restricted conditional sampling and co-occurrence graphs.

Analysis of classifier-free guidance in on-policy diffusion distillation, identifying under-identification issues in velocity matching objectives.

KANEx applies Kolmogorov-Arnold Networks to chest X-ray classification, leveraging spline-based interpretability over black-box vision-language models.

Convergence analysis of Deep Galerkin Method and Physics-Informed Neural Networks for solving nonlinear PDEs.

Multi-turn long-horizon planning for foundation agents via controlled pre/post-training with single/multi-teacher on-policy distillation.

DataOrchestra: per-example curation framework for LLM pretraining that dynamically selects drop/touch/clean operations and processing pipelines.

Claude Opus 4.7 generates shuttling compilers for trapped-ion quantum computers from specifications, handling linear, junction, and general graph architectures.

ERUnderstand: benchmark of 2,960 Entity-Relationship Diagrams for evaluating VLM structured understanding of database schemas.

Fast-slow inference pipelines degrade under deadline pressure; router and merger coordination layers suffer accuracy collapse when slow-path results arrive too late.

Controlled factorial study isolates matcher architecture, model variant, and size effects on entity matching using Qwen3 family across bi-encoder, cross-encoder, and generative approaches.

OpenAI's Hugging Face breach has reignited debate over AI alignment and control, exposing competing views on whether increasingly capable AI should be better aligned, better contained, or both.

Stacked linear combination of unitaries (S-LCU) offers tunable trade-off between barren plateaus and classical simulability in variational quantum circuits.

Co-learning framework for multi-modal classification handles missing modalities at inference via inter-modal collaboration rather than fusion.

Causal-TS: open-source Python library for causal discovery in high-dimensional nonstationary time series with GPU-accelerated conditional independence testing.

Policy distillation framework makes continuous-control DRL interpretable via physics-aware teacher-student pairs; tested on Inverted Pendulum with TD3.

Fixed-lag smoothing reframes bounded inference memory management as estimation; bridges online filters (H=0) and offline optimal eviction strategies.

MMOE applies sparse-expert efficiency principles from LLMs to diffusion transformers, balancing generation quality against training and deployment cost.

APS-RAG: deployed agentic RAG platform at Advanced Photon Source fuses dense, sparse, and knowledge-graph retrieval for institutional knowledge access.

Temporal graph generation models suffer irreducible distribution drift at deployment; masked flow-matching loss decomposition proves observation-based correction impossible.

New behavior capturing a writer's "broad qualities" could have legal implications.

People visit the booth of Kimi, an LLM developed by the Chinese startup Moonshot, during the World AI Conference in Shanghai, China, July 20th. | Image: LONG WEI/ Feature China/Future Publishing via Getty Images Silicon Valley has spent much of the past week on red alert, digesting the arrival of Moonshot AI's Kimi K3, a Chinese AI model that can allegedly beat some of the best systems built by US companies at a fraction of the cost. Its performance alone would have been enough to intensify the rivalry between the US and China. But Moonshot's plan to release the model's weights for free - and...

Kimi K3: 2.8T MoE model with 104B active params, 1M context window, Delta Attention, 2.5x scaling efficiency over K2.

Study on attribution hallucination in VLMs: models answer correctly but fail to identify supporting evidence regions without coordinate supervision.