00:25 — I’ve been tuning into Microsoft’s online Ignite conference. During one of the big keynote sessions, CEO Satya Nadella hosted a conversation with NVIDIA CEO Jensen Huang about how generative AI is the single most significant platform transition in computing history.
01:00 — Acknowledging the AI boom, NVIDIA is providing GPU chips that power the supercomputers that are running AI for Microsoft as well as other cloud providers. Additionally, it’s building hardware needed to run AI.
01:17 — NVIDIA is also using AI themselves. NVIDIA chief scientist Bill Dally recently presented on computer-aided design, demonstrating a new language model that NVIDIA is using internally to assist its chip designers. The new model is called NeMo.
02:14 — Hallucinations can make AI models less useful. A way to overcome this is by training a smaller model with industry-specific data and combining it with the wider knowledge and natural language capabilities of the large model.
02:56 — This process is called retrieval augmented generation (RAG). In this process, the retriever’s job is to search through the smaller dataset that you’ve provided. It’s the generator’s job to produce the final response.
03:12 — The main advantage of RAG is that it allows language models to access and incorporate external information. This expands their knowledge base beyond what they’re trained on, leading to more informed and accurate responses.
03:30 — NVIDIA trained its AI model with its internal data about how its chip works. This enables a chip designer to ask questions and quickly get accurate answers to do their job.
04:07 — Although this language model will not be available outside of the company, I still think it’s interesting to see that NVIDIA is building the chips needed to power AI and also use that same AI to help build more chips.
04:57 — RAGs training a large language model on your smaller specific dataset bridges the gap, giving you the power of natural language processing and general knowledge as well as accurate answers for specific questions.