$ tail -n 3days ./news

Building a Human-Centered AI Future learningmisc

Thinking Machines aims to build AI that extends human judgment through customizable, interactive models, with tools, interfaces, and research that help organizations align AI with their knowledge, values, and local needs.

OpenWiki Brains: Proactive Memory for AI Agents learning

LangChain's OpenWiki Brains gives AI agents proactive memory by gathering and updating relevant context from sources like Gmail, Notion, and Twitter. Version 0.1.0 expands from code documentation to a general-purpose personal brain, using local wikis to keep context current and reduce manual updates across workflows.

Introducing TabFM code

Google's TabFM applies zero-shot learning to tabular data by using dataset examples as context instead of training from scratch. It supports classification and regression, offers a scikit-learn-style interface, and includes open code and pretrained weights.

LLM-as-a-Verifier: A General-Purpose Verification Framework papers

This work presents a general verification framework where an LLM checks if a solution is correct and gives continuous, fine-grained scores instead of only simple judge scores. The method improves verification by using score granularity, repeated evaluation, and breaking criteria into parts, and it reaches strong results on coding, robotics, and medical agent benchmarks. It can also track task progress and provide dense feedback for reinforcement learning.

Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents papers

This work studies long tasks where important earlier information can be forgotten during action, called behavioral state decay. The authors add a separate memory agent that watches recent steps, updates a structured memory, and only gives reminders when needed to the main action agent. It improves pass@1 on two benchmark suites and works better than passive memory or always giving reminders.

Gribouille data viz

Gribouille adds ggplot2-style charting to Typst with layers, scales, facets, smoothers, themes, and an extensive example gallery, making technical publishing more self-contained.

How AI Agents Are Evolving: Skills, Sub-Agents, and Evaluation video

AI agents are shifting from standalone chat interfaces to systems built from reusable skills, specialized sub-agents, and identity-based teammates that operate inside apps, email, and workflows. Practical patterns include model routing for cost control, agentic RAG for unstructured data, ontologies and data agents for structured data, and Work IQ-style access to Microsoft 365. Reliable deployment needs rubric-based evaluation that checks whether an agent followed domain-specific rules, not just whether it completed a task.

GPT-5.6: Frontier intelligence that scales with your ambition news

OpenAI launched GPT-5.6 with three models: Sol, Terra, and Luna. Sol leads in intelligence and efficiency for coding, cybersecurity, and science, with lower token use and cost. The release also adds multi-agent parallel processing, stronger design judgment, and safety measures for advanced research and security work.

GPT-5.6 Series learning

GPT-5.6 Sol, the standout model in the GPT-5.6 series, is the first to win an ARC-AGI-3 public game by correctly reading an unfamiliar scene, using the game's vocabulary, and orienting itself in a new environment.

Evolving Windows vulnerability management to meet the speed of AI-powered discovery learning

Windows uses AI, the MDASH scanning system, and dedicated cloud infrastructure to find vulnerabilities faster, reduce fix times, and deliver timely security updates while maintaining update quality.

Flexible Video Diffusion learning

Flex-Forcing trains video diffusion models to alternate between bidirectional and autoregressive generation with flexible temporal and denoising-step chunking, improving inference speed, video quality, and long-range stability across compute budgets.

Towards Free Normalization: Fusing Normalization into GEMM and Attention Kernels learning

A blog post on kernel fusion techniques for LayerNorm and RMSNorm that reduce memory I/O overhead and speed up memory-bound normalization operations, with code available in the linked repositories.

Muse Spark 1.1 news

Meta released Muse Spark 1.1 with improvements in tool use, coding, computer interaction, and multimodal reasoning, and opened a public preview of its Meta Model API for developers.

Modular Pretraining Enables Access Control learning

Anthropic and AE Studio present GRAM, a modular pretraining method that isolates dual-use knowledge such as virology, cybersecurity, and nuclear physics into switchable modules. The approach lets one pretrained model approximate five separately data-filtered models while using about a fifth of the training compute.

An off switch for dual use knowledge in AI models learning

GRAM lets one model gain the benefits of separately filtered training by isolating dual-use knowledge in dedicated, removable modules. Only those modules are updated on dual-use data, making targeted post-training deletion possible and strengthening model safety.

SWE-1.7: Frontier Intelligence at a Fraction of the Cost learning

Cognition details how SWE-1.7 achieves frontier-level intelligence at lower cost through improvements to its RL pipeline, including infrastructure, training stability, data quality, and long-horizon task techniques. The model is now available in Devin on web, desktop, and CLI.

Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning papers

This work studies asynchronous RL for LLM agents and proposes a single-rollout method, using one rollout per prompt instead of group-wise sampling. It adds value-model training changes and token-level clipping to reduce off-policy problems and keep training stable. The method trains stably for long runs and improves results on coding and reasoning benchmarks, and also works well in simulated online learning where the environment changes over time.