LLMs are typically developed through a process of training on vast amounts of data, the corpus. This costs a lot of time and money. ChatGPT-3, for example, cost $10M. This cost going down but it’s remains expensive. You can avoid this cost for specific use cases by “fine-tuning” a model with specific data or you can augment their prompts with reference data as in Retrieval Augmented Generation or RAG. The next stage in LLM development are models that update/evolve through time. This is what’s discussed in Sakana AI’s paper Transformer²: Self-Adaptive LLMs.
The nice thing about ChatGPT and similar systems is that the complexity of AI/ML functionality is hidden behind a friendly natural language interface. This makes it easily reachable to the masses. But behind this easy to use facade is a lot of advanced functionality that involve a sequence of data processing steps called a pipeline. An AI-powered business card reader, for example, would first detect text and then recognize the individual letters within the context of the words they belong to. A license plate reader would be similar. Detection is an important process that you often need in your AI/ML projects. And that’s why we will be looking at YOLO.
Read More
Leveraging the capabilities of Large Language Models (LLM) using APIs such as the OpenAI APIs is an easy way to add intelligence and advanced functionality to your applications. However, token costs add up and they can get quite expensive. Then there’s the nagging question of privacy and security. Finally, you’re limited in your ability to experiment and customize. But if you have a powerful machine with a GPU or two sitting around, wouldn’t it be great to utilize it for running one of those open source LLMs? Here’s how you can do it.
Read More
Artificial Intelligence (AI) and Machine Learning (ML) require a lot of computing power. Specifically, you would want to have GPUs which have become the standard tool for computing-intensive applications because of their parallel processing capabilities, high throughput, and efficiency in handling the kinds of large-scale computations required for AI and ML.
Read More
The basic artificial neural network we created doesn’t have the best accuracy. Though it performed somewhat well in the recognition of MNIST test data images, it didn’t do as well in the recognition of real-world images. To improve its performance, we can do some adjustments to our model and training data.
Read More