The Archive

Fake citations dashed a dude’s “Are We Dating the Same Guy” revenge lawsuit.

Reddit user reports possible restoration of 60-minute context cache in Claude web interface, contradicting earlier 5-minute reduction.

Anthropic defaults Claude Code's fast mode to Opus 4.7, shifting from prior smaller model.

Dario Amodei projects AI-driven GDP growth alongside 10%+ unemployment, a historically unprecedented combination.

They limited me..but why?

Stainless, a New York-based startup, will wind down all hosted products as part of the acquisition.

Reddit user reports accidental data upload to Claude free tier during work task; raises questions about training data handling and enterprise risk.

llama.cpp MTP speculative decoding merged; Qwen 27B/35B inference speedups 1.24–2.44× on consumer GPUs/APUs with quantization.

The tech trial of the year, Musk v. Altman, was ultimately a fight for control. Elon Musk argued that Sam Altman, with whom he helped found the now-massive company OpenAI, shouldn't direct the future of AI. Altman's lawyers, in turn, poked at Musk's own credibility. A jury came to a verdict on Monday after just two hours of deliberation, dismissing Musk's claims due to the statute of limitations. In a strictly legal sense, three weeks of testimony added up to nothing. But the trial offered a more damning broader takeaway: Almost nobody in this saga seems worth trusting. Some of the most power...

Boston Dynamics Atlas robot demonstrated carrying a refrigerator in a video.

Google DeepMind's Co-Scientist AI identifies genetic factors that reverse cellular aging in human cells.

Musk plans to appeal after judge immediately affirmed the jury's decision.

Elon Musk's lawsuit against Sam Altman and OpenAI dismissed on statute of limitations grounds; Musk plans appeal.

DashAttention enables differentiable, variable-block sparse hierarchical attention via α-entmax, improving gradient flow in long-context LLM inference.

Runtime-Readiness-First Pipeline (RRFP) optimizes pipeline-parallel training scheduling by prioritizing task readiness over static schedules to reduce bubble overhead.

Framework positioning code as unified substrate for agent reasoning, execution, and verification in agentic LLM systems.

ESI-Bench: 10-task benchmark for embodied spatial intelligence via perception-action loops, extending beyond passive observation.

SURGE: derivative-free guidance algorithm for diffusion models using Girsanov reweighting instead of gradient evaluation.

Framework for world models to represent objects as actionable entities with state-dependent properties for physical reasoning.

Vision-OPD: on-policy self-distillation training for MLLMs to improve fine-grained visual detail recognition via regional crops.

Benchmark and attribution method auditing value pluralism in medical LLMs, revealing ethical stance biases in clinical advice.

PIXLRelight: feed-forward single-image relighting via intrinsic conditioning bridging physics-based and learned rendering.

Scaling law: LLM factual recall sigmoid-scales with model size and topic frequency, explaining 60-94% variance across model families.

DexHoldem benchmark evaluates embodied agents on real dexterous manipulation via Texas Hold'em poker with ShadowHand hardware.

General Preference RL unifies online verification-driven and preference-based LLM alignment via multi-dimensional verifiers for open-ended tasks.

Semantic Generative Tuning aligns visual understanding and generation in unified multimodal models via hierarchical generative post-training.

Link: blog.cloudflare.com

Elon Musk's lawsuit against OpenAI dismissed; impacts governance disputes over nonprofit-to-commercial transition.

After around two hours of deliberation, the jury has reached a unanimous verdict in Musk v. Altman, the tech trial of the year. The group found that two claims were barred by the statute of limitations, and a third failed thanks to the dismissal of one of these. The jury here is an advisory jury, meaning the group is installed solely to offer another opinion to the judge, and its verdict is technically not legally binding. Ultimately, US District Judge Yvonne Gonzalez Rogers is the ultimate legal authority - and she accepted the decision. The jury found that Musk's claim for breach of charita...

NDR-SHKF replaces scalar forgetting factors with learned memory attenuation for robust UAV state estimation under sensor noise.