Apr 13, 2025

Google is winning on every AI front

• Google DeepMind's Gemini 2.5 Pro is currently the leading AI model, outperforming competitors like OpenAI and Anthropic across multiple benchmarks, including LMArena and Humanity's Last Exam.

• Google will let companies run Gemini models in their own data centers, enhancing data control and security, with early access slated for the third quarter of 2025.

• OmniSVG is the first end-to-end multimodal SVG generator utilizing pre-trained Vision-Language Models (VLMs), enabling the creation of both simple icons and complex anime characters.

• Skywork-OR1 introduces a series of advanced reasoning models, including Skywork-OR1-Math-7B, optimized for mathematical tasks, achieving a score of 69.8 on AIME24, significantly outperforming similar-sized models.

• Google's decision to allow enterprises to self-host SOTA models marks a significant shift in the landscape of AI, enhancing data privacy and control for businesses, akin to Mistral's existing model.

• d1 is a novel framework that enhances reasoning capabilities in diffusion-based large language models (dLLMs) through a combination of supervised finetuning and a unique critic-free RL algorithm called diffu-GRPO.

• The ML paradox reveals that models excelling in metrics like accuracy can lead to catastrophic real-world failures, such as harmful medical recommendations or erratic self-driving behavior.

• Adding new vocab tokens to LLMs during instruction-tuning has proven ineffective, as models using the base tokenizer show superior validation losses and output quality compared to those with modified tokenizers.

• BSE (Bramble Semantic Engine) is a semantic compressor that efficiently transforms natural inputs into low-dimensional structured representations for language, image, and audio, enhancing preprocessing for LLMs.

• The proposed residual activation function f(x) = x + α · g(sin²(πx / 2)) aims to enhance normalization-free deep networks, potentially improving performance in MLPs and Transformers.

• AI-native runtime enables snapshot-loading of 50+ LLMs (13B–65B) in 2–5 seconds, allowing dynamic execution without constant memory residency.

• ICML 2025: A Shift Toward Correctness Over SOTA?