Making models more efficient is becoming easier in 2024. recent contributions from the open-source community are aligning well with the trend of enhancing the performance of smaller models. This includes not only creating new foundational models but also developing techniques and tools such as open-source data sets for training and fine-tuning models. A standout technique is low-rank adaptation. Instead of directly tweaking billions of model parameters, Laura freezes pre-trained model weights and adds trainable layers to transformer blocks, reducing parameter updates, and thus accelerating fine-tuning with less memory usage. Quantization is also impactful, akin to compressing audio or video files. It lowers data point precision, slashing memory use, and speeding up inference. Merging quantization with low-rank adaptation, known as Chlora, further amplifies efficiency. Direct preference. Optimization gains traction too, offering a simpler alternative to reinforcement learning via human feedback. direct preference optimization strives to align model outputs with human preferences, being lighter and more stable computationally than reinforcement learning from human feedback. These strides, coupled with open source models ranging from 3 to 70 billion parameters, might democratize AI, granting smaller players like startups and enthusiasts access to sophisticated tools once out of reach.