The Archive

US Department of Commerce lifts export controls on Anthropic's Claude Fable 5 and Mythos 5; access restoration begins tomorrow.

Ahmad Osman argues local AI inference is rapidly closing performance/cost gap vs. cloud APIs across consumer and enterprise deployments.

Google releases Gemini 3.1 Flash Lite, optimized for fast, low-cost image generation; author tests visual search capability.

The free open source agentic program is finally invading your phone.

At an event for pharmaceutical executives, biotech founders, and researchers on Tuesday, Anthropic announced Claude Science, a major new product intended to support scientific research in the same way that Claude Code supports software engineering. Like Claude Code, Claude Science can autonomously carry out meaningful work when given concise, high-level instructions, and it has access…

Claude Sonnet 5 launched with performance near Opus 4.8 at lower cost; includes cyber-task restrictions aligned with Opus 4.7/4.8 safeguards.

EquiLibre Technologies, a Prague-based AI lab founded by three ex-DeepMind researchers, is now valued at more than $500 million.

Telling an LLM that 2 + 2 = 5 is enough to make it follow forbidden instructions.

Google's NotebookLM is adding a new way to catch up on your notes: TikTok-style AI videos. The new feature is rolling out to Google AI Ultra and Pro subscribers, allowing NotebookLM to generate 60-second vertical AI clips based on the sources you upload to the app. The example shared by Google details Australia's unsuccessful war on emus, pairing paper cutout-style AI art of emus with narration. It adds to some of the other ways NotebookLM lets you interact with your research, including by generating AI podcasts, cinematic videos, and visual explainers. Doom scrolling but make it educational ...

Google is updating its image generator to make it faster and cheaper, making it a more useful tool for creators looking to make AI content.

They may not look as good, but Nano Banana 2 Lite images only take a few seconds to create.

Nvidia AI chip competitor Etched says it has already booked $1 billion under contract for the inference systems powered by its chip.

Anthropic’s Claude Sonnet 5 brings stronger agentic capabilities, lower pricing, and improved safety, positioning the model as a cheaper alternative to Opus, GPT-5.5, and Gemini Pro.

Anthropic releases Claude Sonnet 5, a frontier model optimized for coding, agents, and professional workflows at scale.

LMs trained on counterfactual explanations from earlier checkpoints produce explanations more faithful to current behavior than source models, advancing mechanistic interpretability.

QVal proposes efficient evaluation method for dense supervision signals in long-horizon LLM agents without expensive end-to-end training.

RL with metacognitive feedback trains LLMs to accurately express uncertainty and improve performance through self-monitoring.

Systematic evaluation reveals data referencing errors (1.7B-20B models misread tables) as overlooked reliability issue beyond final-answer accuracy.

Freeform Preference Learning enables robot policy training from natural-language preference axes instead of binary comparisons.

AdaJEPA adapts latent world models at test time via MPC loop to handle distribution shift in robotic planning.

Generative Skill Composition uses LLM-guided retrieval to select and compose skills for complex agent tasks without exposing full skill library.

Startup Acti is betting the smartphone keyboard is the next home for AI assistants. Its new keyboard for iOS and Android works across apps and lets users create custom AI-powered shortcuts using natural language.

FLORA applies deep learning to LiDAR data for forest attribute prediction under heterogeneous acquisition conditions.

Semantic Reference Frames (SemRF) standardize residual-stream analysis across LM layers to distinguish computation from measurement drift.

Automated Background Swapping reduces classifier reliance on spurious background correlations through synthetic data augmentation.

TRIAGE framework improves credit assignment in agentic RL by classifying action segments into semantic roles, addressing limitations of uniform outcome-based reward.

FedLAB enables federated learning of multimodal graph foundation models with semantic traceability while preserving privacy across decentralized clients.

Skill distillation approach scales browser agent training via behavior cloning from implicit priors in human browsing traces rather than low-level operation.