NVDIA AI some Insights after digesting the event with some Diwali Delicacies!
I had good time at the NVIDIA AI Summit held at the Jio World Center in Mumbai. It felt like every company working on artificial intelligence in India was present, either as an exhibitor or an attendee. Some sessions were so packed that people were standing, and even then, many more were trying to get in.
Much of the discussion revolved around Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG). These are undoubtedly central topics in AI today, drawing significant attention. I’m a huge admirer of RAG, but I still wouldn’t go as far as to say that “LLM (+RAG) is AI.” Although no one at the conference explicitly said this, it felt implied over the three days of sessions.
I may be wrong, but I sensed a push for these technologies. Not just from NVIDIA, but from any hardware supplier, there’s an incentive to promote anything that drives demand for their solutions. NVIDIA’s GPUs are a backbone of recent AI advancements, so it’s understandable that they’re pushing hard to maximize the potential of this technology. Hence, the emphasis on encouraging everyone to create or fine-tune their own LLMs and develop their own AI. Again, I could be mistaken, but this is my personal view.
Let me explain it with an analogy. Imagine you're in a suburb with a 10 km radius—say, Kharghar, where our office is located. Now, imagine you’re asked to explore every possible road in Kharghar, with each iteration taking a different route. Eventually, given a limited number of roads, you'll end up with a comprehensive list, say Ra, Rb, Rc, and so on. If another person does the same exercise, they’ll likely come up with the same list, albeit perhaps in a different order.
This analogy illustrates my view on training data for LLMs. Most models have access to the same “open” data, and legal cases are even revealing more proprietary data. When two LLMs are trained on the same data, likely using the same transformer-based technology with minor tweaks, they won’t learn or generate anything fundamentally different. The real issue often lies in the prompt, not the model itself. As people say, some things that should just be a feature in an existing product shouldn’t be turned into a separate startup. Similarly, what could be achieved with a well-constructed prompt doesn’t necessarily warrant a new LLM or even fine-tuning an existing one.
On the other hand, AI isn’t just LLMs with RAG. Recently, I spoke with some final-year engineering students specializing in data science, and I asked them how they’d build a classifier for English language helping verbs. They gave answers ranging from BERT-based approaches to RNNs and even Random Forest classifiers. I know it’s not entirely their fault—this is the type of AI that’s being celebrated right now.
It’s like the California Gold Rush. Back then, it was Levi Strauss of Levi’s jeans who thrived, and now it’s Jensen Huang of NVIDIA. Yes, in both the cases there is indeed “gold”, but there’s also a RUSH, and NVIDIA stands to gain a lot from it. But we also need to understand some of it’s a win-win for stakeholders with who can gain significantly out of it. However, we must do our homework before blindly following any advice.
We have a few A100s and certainly appreciate the hardware, but it’s worth questioning: do you truly need the machine, or do you truly need to train or fine-tune your own LLM Or there are alternatives?
My take away from the conference :
There is really Gold🪙, but there is a rush too.
Participate wisely!
Comments