AI Edition - It's AI Time - 336 [17-03-2025]

AI Edition - It's AI Time - 336

Google Assistant is Replaced by Gemini:

Google Assistant is transitioning into Gemini, a more personalized and AI-driven assistant designed to integrate with apps and services while leveraging generative AI.

Google's Response to U.S. AI Policy:

Google outlines its vision for U.S. AI policy, advocating for investments in AI infrastructure, streamlined government adoption, and international pro-innovation standards to maintain leadership in the AI space.

Sakana claims its AI-generated paper passed peer review:

Japanese AI startup Sakana used its AI Scientist-v2 to generate a paper accepted at an ICLR workshop but withdrew its post-initial review due to transparency concerns. Though Sakana highlights AI's potential, critics argue that human judgment was crucial and acceptance was facilitated by higher workshop acceptance rates. Skeptics remain wary, emphasizing AI's technical flaws and the need for established norms to prevent degrading the scientific review process.

Luma's new pre-training method for multi-modal models:

Luma Chief Scientist Jiaming Song, who created the earliest accelerated algorithm for diffusion models, has unveiled Inductive Moment Matching (IMM) a new multi-modal pre-training method that surpasses diffusion models with superior sample quality and 10x greater efficiency.

ThunderKittens on Blackwell:

ThunderKittens is a powerful and simple abstraction for writing efficient and performance CUDA kernels. This post explores how to use the framework for the newest Blackwell series of GPUs from Nvidia. The main difference is thinking in terms of data flow.

Are Open Models Really Open?:

Many AI models labeled as "open" come with restrictive licensing terms. Google's new Gemma 3 models and similar releases from Meta raise concerns about commercial limitations, potentially impacting smaller companies relying on these technologies.