Jul 5, 2025

• ChatGPT creates phisher's paradise by serving the wrong URLs for major companies
  AI chatbots, like ChatGPT, incorrectly provide URLs for major companies only 66% of the time, creating a fertile ground for phishing attacks. This vulnerability arises as criminals exploit the AI's tendency to generate misleading links, potentially leading users to fraudulent sites.

• A new, 200% faster DeepSeek R1-0528 variant appears from German lab
  TNG Technology Consulting GmbH's new DeepSeek-TNG R1T2 Chimera model is 200% faster than its predecessor, R1-0528, achieving 90-92% of its intelligence while using 40% fewer output tokens, thus enhancing efficiency and reducing costs.

• Just Ask for Generalization
  Generalization may be more effective than direct optimization in building AI systems, as large, diverse datasets enhance model performance more than intricate biases, aligning with findings from Neural Scaling Laws and Foundation Models.

• What I learned building an AI coding agent for a year
  Building an AI coding agent revealed that initial optimism masked significant reliability issues, with only 5-10% of tasks failing despite early growth to $5k MRR during YC.

• Self-Correction Blind Spot: Revealing and Addressing the Self-Correction Blind Spot in LLMs
  LLMs struggle to correct their own outputs, with a 64.5% average blind spot rate across 14 models, indicating a significant gap in their self-correction capabilities.

• kappaTune: a PyTorch-based optimizer wrapper for continual learning via selective fine-tuning
  kappaTune is a PyTorch-based optimizer wrapper that selectively fine-tunes model parameters based on their condition number (κ), focusing on the least anisotropic parameters to maintain pre-trained knowledge and reduce catastrophic forgetting.

• An analytic theory of creativity in convolutional diffusion models.
  This study presents an analytic theory of creativity in convolutional diffusion models, revealing that locality and equivariance induce combinatorial creativity, allowing models to generate original images beyond mere memorization of training data.