The AI landscape has been transformed dramatically in recent months, thanks to the release of LLaMA (Language Learner and Model Aggregator) by Meta. With its open-source code and subsequent public leak, the AI community has been witnessing an unprecedented acceleration in innovation, collaboration, and accessibility to powerful language models. This blog post will delve into the critical milestones in the LLaMA revolution, shedding light on the democratization of AI and its implications for the future.
I. LLaMA: A Catalyst for Change Meta launched LLaMA on February 24, 2023, open-sourcing the code but not the weights. Although the model was relatively small (available at 7B, 13B, 33B, and 65B parameters) and not instruction or conversation tuned, it was quite capable relative to its size. The public leak of LLaMA on March 3, 2023, opened the floodgates for experimentation and innovation, profoundly impacting the AI community.
II. Key Milestones in the LLaMA Revolution
Language Models on a Raspberry Pi: On March 12, 2023, Artem Andreenko got LLaMA working on a Raspberry Pi, setting the stage for minification efforts.
Fine-Tuning on a Laptop: On March 13, 2023, Stanford released Alpaca, which added instruction tuning to LLaMA. Eric Wang's alpaca-lora repo enabled low rank fine-tuning, making it possible for anyone to fine-tune the model on a single RTX 4090.
Practicality on a MacBook CPU: Georgi Gerganov's 4-bit quantization made LLaMA run on a MacBook CPU, providing the first practical "no GPU" solution.
Vicuna: A 13B model achieving "parity" with Bard for just $300 in training costs, circumventing restrictions on ChatGPT API.
GPT4All by Nomic: An ecosystem for gathering various models (including Vicuna) in one place, with a training cost of $100.
Open Source GPT-3 by Cerebras: Outperforming existing GPT-3 clones, this marked the first confirmed use of μ-parameterization "in the wild."
LLaMA-Adapter: Achieved instruction tuning and multimodality in one hour of training, with just 1.2M learnable parameters, setting a new SOTA on multimodal ScienceQA.
Koala: A Berkeley dialogue model with human preferences comparable to ChatGPT, with a training cost of $100.
III. The Impact and Future of Democratized AI The LLaMA revolution has made powerful AI models accessible to a broader audience, fueling innovation and collaboration. This democratization of AI is undoubtedly exciting, as it holds the potential to drive groundbreaking advancements in various fields.
However, this rapid democratization also raises questions about ethical and safety concerns. As more people gain access to powerful AI tools, it becomes crucial to ensure responsible development and use of these technologies.
Conclusion: The LLaMA revolution has reshaped the AI landscape, creating new opportunities for innovation and collaboration. As we embrace this new era of open-source AI, we must collectively work to address the challenges it poses and ensure the responsible development and use of these powerful tools. The LLaMA revolution demonstrates that the future of AI is open, collaborative, and full of potential for transforming our world.