If OpenAI and Google are focused on creating “models,” then Hazy Research is working on the “infrastructure behind the models.” Transforming the underlying computation of AI to enable different chips to efficiently run large models is pushing AI from being “hardware-constrained” to “hardware serving software”.
Image source: WikipediaHazy Research is a top AI research team at Stanford University, led by computer science professor Chris Ré. Why is this team gaining attention? Because they are not conducting “paper-level research,” but rather developing foundational technologies that can truly change the entire AI engineering system. You may have heard of their achievements, including:1. Snorkel: The pioneer of weak supervision learning (making AI no longer reliant on manual labeling)2. Overton: Automatically building machine learning systems (automating the generation of AI systems)3. MosaicML’s early concepts laid the groundwork (making large model training cheaper)4. Karpathy (former head of AI at OpenAI/Tesla) also mentioned their influenceIn 2024-2025, they are undertaking a project of significant industry importance: how to free AI from being “hardware-bound” so that various accelerator cards (AMD, NVIDIA, self-developed chips) can truly run faster.Why is this so important? It can be viewed from the following three points: They continuously release:1. ThunderKittens (TK): A high-performance CUDA Kernel framework2. HipKittens (HK): A high-performance Kernel framework for AMD GPUs3. AI compilers, tile abstraction, cross-chip scheduling and other core technologies that may be reshaping the entire AI hardware ecosystem. In simple terms: Hazy Research is not optimizing models, but rather “inventing methods to make AI faster, cheaper, and more scalable— optimizing the underlying infrastructure for AI operations.📍In the next issue, we will detail 【HipKittens: The “Secret Weapon” That Truly Makes AMD GPUs Run】

Click the blue text
Follow us to learn more about cutting-edge AI technology~