Feb 17, 2026

• Claude Sonnet 4.6 is a significant upgrade, featuring enhanced skills in coding, long-context reasoning, and a 1M token context window, making it the most capable model yet from Anthropic.

• Tesla's "Robotaxi" fleet in Austin has reported 14 crashes since its launch, with five new incidents in just one month, indicating a troubling trend of safety concerns. The crash rate now stands at one crash every 57,000 miles, significantly worse than the average human driver, who experiences a minor collision every 229,000 miles.

• Qwen operates across multiple domains, including qwen.ai and chat.qwen.ai, indicating a robust infrastructure for AI interactions and services. Link to article

• SkillsBench evaluates 86 tasks across 11 domains, revealing that curated Skills improve average pass rates by 16.2 percentage points, with significant variability across domains, from +4.5pp in Software Engineering to +51.9pp in Healthcare.

• Chess engines can achieve superior performance through distillation from search, allowing a bad model combined with effective search to outperform a good model alone, as evidenced by a ~200 elo difference in model quality.

• Over 18,000 OpenClaw instances are publicly exposed, with nearly 15% of community skills harboring malicious instructions that can download malware or exfiltrate sensitive data.

• Molecular networks exhibit agency, suggesting that even simple biomolecules may possess goals and desires, potentially reshaping our understanding of life and consciousness.

• Elephant trunk whiskers possess a unique stiffness gradient that enhances their tactile sensing capabilities, allowing elephants to discern contact points along their 1000 whiskers, crucial for delicate tasks like picking up food without damage.

• Linear RNNs demonstrate superior performance in state-tracking tasks, particularly in permutation composition, by converting these tasks into code via REPL traces, which reveal states through prints and variable transformations.

• SparseFormer addresses the O(n²) compute bottleneck in vision transformers, showcasing significant potential for commercial applications like data labeling and industrial inspection.

• The NVIDIA Blackwell Ultra platform achieves up to 50x better performance and 35x lower costs for agentic AI, driven by advanced hardware and software optimizations that enhance throughput and efficiency.

• MicroGPT-C is a zero-dependency, pure C99 implementation of a GPT-style model, achieving dramatically faster training and inference compared to Python, with training times reduced to 20 ms and inference to microseconds.

• ReusStdFlow introduces a novel "Extraction-Storage-Construction" paradigm to tackle the reusability dilemma in enterprise Agentic AI, enabling the deconstruction of platform-specific Domain Specific Languages (DSLs) into standardized workflow segments.

• BFS-PO is a novel RL algorithm that employs a Best-First Search strategy to mitigate the overthinking problem in Large Reasoning Models (LRMs) by focusing on generating the shortest correct answers through a backtracking mechanism based on maximum entropy nodes.

• Hallucinations in small-sized language models (LLMs) are characterized by fluent yet factually incorrect responses, which can be analyzed through a geometric lens, revealing that genuine responses cluster more tightly in the embedding space.