Five Lies About AI

Everyone is talking about AI lately! Which of the things you hear may not exactly be true or in some cases be outright incorrect? Let’s take a look at 5 inaccuracies that are rather popular.

More Data is Better for Training and Fine Tuning Models

Not really! More of the same kind may actually make the model performance worse, while it drains your pocket. Variety, variance, and entropy in the data tend to be more helpful. Intuitively, we learn more when we encounter many different types of situations.

Poor Data Quality & Lack of Structure Is Hindering AI

In reality, data inconsistency and more unstructured data, as opposed to structured data, is where language models and NLP at large shines. If you have clean and structured data, you may see traditional analytics and machine learning models give more value with higher accuracy as compared to an LLM powered alternative.

Hallucination is a Major Problem in LLMs

LLMs are probabilistic machines that predict the next token. Fluency is its strength and not its weakness. One could argue that hallucination is a feature and not a bug. Perhaps you are using this power tool for the wrong job?

Nvidia GPUs outperform all Accelerators

Nvidia GPUs are great and Nvidia is the darling of this AI surge. However, the core advantage it derives is from its extremely robust software layer, CUDA, and the other middleware components. It can easily run most deep learning models, large language models, and a variety of other machine learning algorithms and systems. TPUs from Google and GPUs from AMD are not worse, they just lack the software layer versatility that Nvidia has.

Best Performing LLM will Rule It All

While much debate, benchmarking, and conversation is around LLM model performance, most real world successful AI systems use multiple techniques and often multiple models to achieve superior performance. Perhaps, it makes sense! An uber smart jack of all trades is likely to get outcompeted on a specialized task against someone who has expert level skills in doing that particular task well.