Albert Gu

Albert Gu, assistant professor of machine learning at Carnegie Mellon University, is working on giving artificial intelligence something akin to memory.

Currently, every time you ask models like ChatGPT a question, it considers every prior bit of information the user has provided before generating a response. This is the way most large language models (LLMs) work and contributes to the lag users experience when chatting with the AI. 

Gu, who is also a co-founder of the AI startup Cartesia, has developed a new way of designing models that allows the AI to compress every prior data point into a “summary of everything” it has seen before, Gu describes. In a paper published in December, Gu introduced this design, called “Mamba,” which gives models something like a working memory. This potentially makes them faster than traditional LLMs, and much more efficient in how they draw on computing power, particularly in domains beyond language, such as audio and genetics. The approach seems to be taking off: Abu Dhabi's Technology Innovation Institute (TII) has implemented Gu's Mamba architecture in their Falcon Mamba 7B model, creating one of the world's first open-source models to use this technology. 

Gu, who has been working on this approach for almost five years alongside his colleague Princeton University professor Tri Dao, and others, emphasizes that the design approach represented by Mamba offers a “different paradigm of working with data.” Like many in the field, his ultimate goal is a system general enough to handle anything.

Gu believes his work at Carnegie and Cartesia shows there are many paths development of artificial intelligence could follow.  “Even just the proof of concept that there are companies who have scaled these models—really good models—based on these alternate architectures offers an important signal to open source and academic communities,” Gu says