The Archive

Anthropic Frontier Red Team finds GLM-5.3 and Claude Mythos Preview achieve control flow hijacks on 4-6% of binary exploitation tasks; earlier models failed entirely.

New sculpture depicts AI leaders escaping a sinking ship.

Elon Musk | Image: Laura Normand / The Verge Grokipedia, the AI-powered online encyclopedia from SpaceXAI, appears to be updating articles once again after a months-long pause. In August, Lawfare reported that articles on Grokipedia hadn't reviewed edits since April, but the platform's live updates site is now showing various recent changes to pages - though as I write this, many are just a note that says "recheck all references and sources." The site seems to have become active again relatively recently. The page for President Barack Obama has a note that says it was "Fact-checked by Grok" t...

The deal, which is expected to close by year's end, is worth $8.2 billion.

OpenAI is building out the pieces of an alternative to the traditional app store model, turning ChatGPT into a place where software can be discovered and used by people and AI agents alike.

The new round is anticipated to be the company's last before its delayed 2027 public debut.

Parallel work, model-family isolation, reversible changes, and GPU-backed validation shaped an open source project designed around coding agents NVIDIA TensorRT... Parallel work, model-family isolation, reversible changes, and GPU-backed validation shaped an open source project designed around coding agents NVIDIA TensorRT Model Connect is an open source collection of AI model reference implementations in C++, built on top of NVIDIA TensorRT. It began with a practical question: could the performance of the NVIDIA inference stack be made accessible to… Source

Vision-language models have made it possible to build visual AI agents that understand video at production scale. The harder problem is turning that capability... Vision-language models have made it possible to build visual AI agents that understand video at production scale. The harder problem is turning that capability into a maintainable system that combines ingestion, stream processing, event detection, retrieval, summarization, and reporting. The NVIDIA Metropolis Blueprint for Video Search and Summarization (VSS) and its agent skills help… Source

OpenAI isn't a public supporter of Nvidia's Open Agent Safety Platform, but it is privately working with Nvidia, TechCrunch has learned.

Simon Willison's brief commentary on GPT-6.1 Sol pricing vs. performance; notes similarity to GPT-6 family in capability visualization.

Without a "full set of safeguards," agent accessed "system information and source code."

Novel framework for autonomous robot policy improvement using foundation models to compose learned behaviors and self-correct from failures without human demonstration.

Technique enabling Multimodal LLMs to construct 3D scene representations from multi-view images before answering visual reasoning questions, improving cross-viewpoint coherence.

Theoretical analysis proving relationship between semantic speciation and nonlocality windows in diffusion model dynamics via spatial semantic information distribution.

Method for quantizing recurrent states in linear attention models by targeting temporal persistence and spatial row-specific error impact, reducing memory bottlenecks.

Quantization technique for hybrid LLMs with linear attention (Gated DeltaNet, Kimi Delta Attention) addressing outlier rows/columns and rounding error accumulation in recurrent states.

Hybrid transformer-CNN model (PAtteRNS) applying separate self-attention to temporal, spectral, and spatial dimensions for crop segmentation in Sentinel-2 satellite imagery.

Spectral theory proving sharp bounds on Wasserstein distance between original and LLM-reconstructed graphs, improving graph reconstruction evaluation methodology.

Training-free vector steering method (EmoRES-TTS) enhancing emotion controllability in frozen TTS models through residual-enhanced emotional representation modification.

Framework (GEB) for long-video QA grounding visual entity observations across hours via biographies, resolving identity ambiguity across disjoint events.

LIFT architecture enabling recurrent state propagation across transformer layers during pretraining, eliminating feed-forward-only bottleneck for information reuse during generation.

Meta-reasoning harness for agentic inference: controller orchestrates worker tasks, resource budgets, and execution flow for long-horizon problems.

Test-time AI4AI: Builder learns frozen Meta-Skills to design execution harnesses for fixed Target models across novel tasks.

Small trainable advisor guides frozen frontier LLMs via natural-language guidance, learning from task rewards and interaction feedback.

SplitMoE applies spatiotemporal-aware Mixture-of-Experts routing to video diffusion, breaking uniform expert-load regularization.

LongHarness Bench evaluates long-context LM harnesses on retrieval + reasoning tasks requiring diverse strategies over extended contexts.

DeGG-Flow: flow matching framework for multi-agent generation with decoupled guidance ensuring hard constraint satisfaction.

Theoretical bound O(1/N) optimality gap for weakly-coupled MDPs with budget constraints, generalizing restless bandit results.

WUSH-KV applies data-adaptive transforms to KV-cache quantization, reducing memory/bandwidth for long-context inference.

Stochastic world models as verifiable perception surrogates for vision-based neural feedback systems, replacing GANs.