Mar 5, 2026

Something is afoot in the land of Qwen

• Qwen 3.5 models, recently released by Alibaba, showcase impressive capabilities, yet the team faces uncertainty following the resignation of lead researcher Junyang Lin, a pivotal figure in their development.

• Fine-tune Qwen3.5 models locally with Unsloth, achieving 1.5× faster training and 50% less VRAM usage compared to FA2 setups, with support for both vision and text fine-tuning.

• Relicensing challenges in open source are exemplified by the chardet project, which transitioned from LGPL to MIT using AI-assisted rewriting, raising questions about copyright compliance and derivative works.

• Huginn identified 254 confirmed phishing sites in February 2026, revealing that 83.9% were missed by Google Safe Browsing, highlighting the limitations of reactive detection methods in combating rapidly evolving phishing tactics.

• Clinejection exploited a prompt injection in a GitHub issue title, leading to the silent installation of OpenClaw on 4,000 developer machines through a compromised npm package.

• NanoGPT Slowrun has achieved 5.5x data efficiency in just a week, significantly enhancing learning algorithms for language models by leveraging unlimited compute with limited data.

• The d² Pullback Theorem posits that the essence of Attention is a d²-dimensional problem, challenging the prevailing belief that it is n²-dimensional, revealing a fundamental misunderstanding of its intrinsic geometry.

• AI's ability to re-implement code is transforming software development, as demonstrated by an AI porting a library to a new language with a different design while maintaining similar functionality, highlighting the evolving landscape of coding practices.

• Generative Flow Networks (GFlowNets) significantly enhance radio propagation modeling by achieving speedups of up to 10x on GPU and 1000x on CPU, while ensuring high coverage accuracy through intelligent path sampling.

• ORION is the first open-source system enabling native training of a 110M Transformer on the Apple Neural Engine (ANE), overcoming limitations imposed by CoreML's opaque abstractions and lack of on-device training support.

• Aura-State is an open-source Python framework that compiles LLM workflows into formally verified state machines, enhancing reliability by integrating techniques from hardware verification and statistical learning.

• Phi-4-reasoning-vision-15B is a 15 billion parameter multimodal reasoning model that excels in math and science reasoning, offering efficient performance across various vision-language tasks while minimizing compute costs.

• GLiNER2 integrates Named Entity Recognition, Text Classification, Structured Data Extraction, and Relation Extraction into a single 205M parameter model, enabling efficient processing without external dependencies.

• Open-sourced framework enables the creation of physics-simulated humanoids in Unity, utilizing MuJoCo for realistic dynamics and allowing on-device reinforcement learning (RL) training directly in Unity without external servers.

• The EU AI Act imposes stringent compliance standards on high-risk models, such as those used in credit scoring and insurance pricing, significantly affecting how data science practitioners develop and maintain these systems.