Anthropic is investing CAD 10 million in 8 Canadian institutions with a focus on AI safety, healthcare, and low-resource languages. A simultaneously published Economic Index reveals that Canada accounts for 2.6% of global Claude.ai traffic, with per-capita adoption 4.4 times higher than expected.
Anthropic has launched a free version of Claude for verified K-12 teachers in the US, available until June 2027. The program includes a library of teaching skills, a curriculum aligned with educational standards from all 50 states, and integrations with 9 educational tools.
Test-time compute is the additional computation a model spends during inference to produce a better answer. Bojie Li and Noah Shi define interaction with external tools as the third axis of test-time compute — alongside longer thinking and best-of-N sampling. A proposer-reviewer system achieves a 100% pass rate, while self-thinking and best-of-N plateau.
SCALECUA is a new framework from Tsinghua/THUDM researchers that scales computer-use agents using online reinforcement learning, achieving a new SOTA result of 68.7% on the OSWorld benchmark and 54.0% on ScienceBoard.
Claude Code v2.1.208 introduces three new capabilities: an opt-in screen reader mode for accessibility, vim insert-mode key remapping, and a corporate process wrapper covering all self-spawned processes. Hotfix v2.1.209 immediately fixed a bug that was blocking dialogs in background agent sessions.
GitHub has introduced AI-generated security detections directly on pull requests in Code Scanning, along with a new /security-review command in the Copilot app that analyzes code for injection, XSS, path traversal, and weak cryptography.
NVIDIA Blackwell GB300 NVL72 is up to 25× more efficient than Hopper measured by performance per watt on the DeepSeek V4 Pro model, and software-only optimizations within a single month delivered up to a 5× improvement.
AMD has released LogsLop, an open-source Python CLI tool that compresses large log files through normalization and clustering. On 73 real log files it reduces the line count to 11% and bytes to 7.5%, and with Llama 3.3 70B it uses 78% fewer tokens.
A terminal agent is an AI agent that solves tasks through commands in a terminal or shell. SETA (Shen et al., Oxford, KAUST) generates 4,500+ verifiable RL environments for training terminal agents. DeepSeek-V4-Flash improves pass@1 from 40% to 43% and pass@5 from 54% to 58% on Terminal-Bench 2.0.
Autoformalization is the automatic translation of natural-language mathematics into a machine-verifiable formal form. The ToMap pipeline by Tian-Shuo Liu and collaborators achieves a 19.0% improvement over the previous SOTA method on the ProofFlowBench dataset. The key innovation is concentrating test-time compute on the Decomposer as the bottleneck instead of distributing it evenly.
Multi-agent orchestration is the coordination of multiple specialized AI agents on a shared task. Thrad.ai implemented a system for finding B2B buyers across 6 sources using Amazon Bedrock and the Strands Agents SDK. The Swarm pattern achieves email quality of 8.2/10, while the Graph pattern costs 25% less and runs 28% faster.
Lin Sun from Solo.io questions whether the Kubernetes Pod is the right abstraction for deploying AI agents. Agents are ephemeral — they activate briefly, and a dedicated idle Pod wastes resources. He proposes a new 'agent-substrate' model with Actors and WorkerPools.
Google DeepMind and the Atal Innovation Mission have launched ATL Saathi, a Gemini-powered web application that helps teachers in Indian school STEM labs. It is piloting in 100 schools and targets more than 11 million students, with support for 8 languages and 12 teaching modules.
Allen Institute for AI published a detailed analysis of the maritime agent Shippy's architecture, which serves 70+ countries on Claude Opus 4.6. The key takeaway: agent reliability depends less on model strength and more on deterministic tools, isolated infrastructure, and evaluations grounded in real operational workflows.
AMD's GEAK agent for automated GPU kernel optimization migrated the DeepSeek-V4 MLA kernel from PyTorch to Triton for MI355 accelerators. Results show up to 9.13× prefill speedup, 4.94× geomean decode speedup, and 2.10× higher end-to-end throughput in the SGLang framework.
Anthropic analyzed 309,815 anonymized Claude.ai conversations and found statistically significant differences in expressed values between models Sonnet 4.6, Opus 4.6, and Opus 4.7, as well as across 20 analyzed languages — Hindi shows the highest warmth, while Russian and English show the highest rigor.
Li Auto, a Chinese automaker, has released its own foundation model Mach-Mind-4-Flash. It is a 35-billion parameter MoE model that activates only 3 billion parameters per input, and through a three-stage training pipeline achieves performance in the 100B+ model class with dramatically lower inference costs.
A new architectural framework for agentic LLM systems introduces selective persistent memory that retains four categories of reusable context while discarding transient reasoning traces. The result is 96% task completion success and 97× lower token cost per call.
Soofi S is a 30B MoE hybrid Mamba-Transformer model with 3B active parameters pre-trained on approximately 27 trillion tokens, including enhanced German language data. Trained exclusively on Deutsche Telekom's cloud infrastructure, it outperforms all European sovereign alternatives and leads the code benchmark among 17 open models.
'From Solvers to Research' is an arXiv paper, co-authored by Fields Medal recipient Terence Tao, examining whether LLM agents can perform research-level mathematics through auto-formalization and proof synthesis. The paper identifies systematic weaknesses in existing proof-synthesis systems and maps the boundary between solving problems and genuine research.
Token-Flow Firewall is a defensive mechanism presented on arXiv that monitors token flows in natural language in real time for persistent AI agents. In tests against prompt injection and other adversarial attacks it reduces the attack success rate to 12.5%, offering a runtime protection layer instead of relying solely on model training.
Claude Code v2.1.207 is a new release of Anthropic's CLI tool that makes 'auto mode' automatically available on Amazon Bedrock, Google Vertex AI, and Microsoft Foundry without an opt-in variable. The release closes a shell-injection security vulnerability in hook configuration and switches the default model to Claude Opus 4.8 on Bedrock, Vertex, and the Claude platform on AWS.
CausalDS is a benchmark presented on arXiv that tests whether AI agents can distinguish causality from correlation within data-science workflows. It fills a gap in the evaluation of autonomous analytical agents, where standard tests measure execution accuracy but not the ability to reason correctly about causation.
Overthinking is a paper accepted at ICML 2026 showing that amplifying the reasoning weight in large language models can extract hidden learned information the model would not otherwise expose. The finding opens a new class of extraction attacks targeting reasoning models such as o1, DeepSeek R1, and Claude with extended thinking.
Super Weights in LLMs is a paper accepted at COLM 2026 that identifies 'super weights' — a small number of parameters whose modification disproportionately changes a language model's behavior. The paper shows that these weights explain why selective fine-tuning of only certain layers often fails, with direct implications for PEFT and LoRA approaches.
Amazon SageMaker AI has introduced serverless customization for NVIDIA Nemotron 3 models, requiring no infrastructure management. Three techniques are available: SFT (supervised fine-tuning), RLVR (reinforcement learning with verifiable rewards), and RLAIF (reinforcement learning from AI feedback), making advanced RL methods accessible to enterprise teams without ML infrastructure expertise.
CodeQL 2.26.0 is the new version of GitHub's static security analysis tool, introducing detection of AI prompt injection attacks as a new analysis type, along with support for Kotlin 2.4.0. It is the first integration of an AI-specific attack vector into a mainstream SAST tool, bringing prompt injection into the same security workflows as XSS and SQL injection.
SensorFM is Google's foundation model for health data from wearable devices, trained on more than one trillion minutes of signals from Fitbit and Pixel Watch devices worn by 5 million users in over 100 countries. The model outperforms specialized approaches on 34 of 35 tasks, with +9% AUC on classification and +21% correlation on regression.
Aurora 1.5 is Microsoft's open-source foundation model for the Earth system that outperforms the ECMWF ensemble forecast on 88.9% of evaluated variables and horizons. The new release adds 22 meteorological variables, hourly resolution, and probabilistic ensemble forecasting — achieving around 33% lower track error for Hurricane Helene compared to the original Aurora.
GPT-5.6 is OpenAI's new model family with three variants: Sol (flagship for complex reasoning), Terra (balanced), and Luna (high-volume, cost-efficient). It introduces Programmatic Tool Calling, explicit prompt cache control, persisted reasoning and multi-agent orchestration in beta — available from day one in GitHub Copilot.
Ben Bernanke is the former chair of the US Federal Reserve (2006–2014) and 2022 Nobel laureate in economics, appointed by Anthropic to its Long-Term Benefit Trust. The LTBT is an independent body with no equity or profit interest that oversees the company's public benefit mission — Bernanke brings expertise on the economic impacts of AI.
Mistral introduced Robostral Navigate, its first model for embodied robotic navigation with 8 billion parameters. Using only a single RGB camera — no LiDAR or depth sensors — it achieves 76.6% success on the R2R-CE benchmark for unseen environments and surpasses multi-sensor competitors by 4.5 percentage points.
OpenAI introduced GPT-Live, a new voice model designed for natural, lifelike conversational AI interactions. Integrated into ChatGPT Voice from day one, it arrives as part of an accelerated voice model development cycle — just two days after GPT-Realtime-2.1. Technical specifications and pricing have not yet been published.
PyTorch 2.13 ships with 3,328 commits from 526 contributors. Key highlights: nn.LinearCrossEntropyLoss cuts peak GPU memory footprint by up to 4× for LLM training, FlexAttention on Apple Silicon achieves up to 12.3× speedup, and the new torchcomms backend modernizes distributed training.
Anthropic and AE Studio announce GRAM (Gradient-Routed Auxiliary Modules) — a method that isolates dual-use knowledge such as virology, cybersecurity, and nuclear physics into removable neural modules during training, enabling a single training run to produce multiple model variants with different capability sets.
The European Commission today presented a joint Action Plan on Cybersecurity and Artificial Intelligence that builds on the existing EU legal framework, coordinates member states and industry, and addresses AI simultaneously as an attack vector and a defensive tool.
Meta Superintelligence Labs has unveiled Muse Image and Muse Video — models that operate as agents, internally invoking code and web-search tools, ranking #2 and #3 on the Arena leaderboard, with mandatory Content Seal watermarking.
The UK AI Security Institute used frontier models to audit its own research platform and found a five-step attack chain that had previously escaped standard tools — all for under £150 in token costs.
Researchers from Seoul National University and the Universities of Indiana and Wisconsin introduce Agent Data Injection (ADI) — a new attack type that embeds malicious data into trusted data structures of agents, achieving arbitrary click attacks and remote code execution on Claude Code, Codex, and Gemini CLI while bypassing existing defenses.