TL;DR:
- Open source developers may outshine tech giants like Google and OpenAI in the generative AI market.
- Running large language models (LLMs) on consumer hardware is now feasible, enabling personalized AI on smartphones and laptops.
- Meta’s leaked Large Language Model Meta AI model (LLaMA) has sparked innovation in the open source community.
- Low-rank adaptation (LoRA) has revolutionized fine-tuning, reducing the cost and time required.
- Open source AI models offer attractive alternatives due to rapid innovation and lack of usage restrictions.
- The success of FAANG companies in generative AI is built upon open source AI programs.
- Meta has the potential to leverage its code for future products, and other companies might benefit from collaboration with open source developers.
- Open source has significant advantages that make it a formidable force in the market.
Main AI News:
A leaked memo from Google has brought to light a surprising revelation: open source developers may be the ultimate winners in the generative AI battle, overshadowing tech giants like Microsoft and Google themselves. In this rapidly evolving landscape, a hidden third party has quietly emerged as a formidable force, and it’s none other than the open source community.
One might wonder, how is this possible? Isn’t the delivery of large language models (LLMs) and high-quality answers reliant on hyperscale clouds, such as those provided by Google and OpenAI? Well, the truth is, the game has changed. Running LLMs on a smartphone is now a reality, with foundation models effectively operating on a Pixel 6 at an impressive rate of five LLM tokens per second. Furthermore, it has been demonstrated that personalized AI can be fine-tuned on a laptop in just a few hours.
The ability to personalize language models quickly and efficiently on consumer hardware marks a significant milestone in the field. As one expert puts it, “Being able to personalize a language model in a few hours on consumer hardware is a big deal, particularly for aspirations that involve incorporating new and diverse knowledge in near real-time.“
So, what triggered this revolution? The answer lies in Meta’s Large Language Model Meta AI model (LLaMA), which was recently leaked. This event served as a catalyst, sparking a wave of innovation within the open source community. Despite lacking initial instruction or conversation tuning, the model underwent rapid iterations, resulting in remarkable enhancements such as instruction tuning, quantization, quality improvements, and more, all achieved in quick succession.
Among the groundbreaking developments, one technology stands out: low-rank adaptation (LoRA). This cost-effective fine-tuning mechanism has significantly lowered the barrier to entry for training and experimentation. As a result, individuals can now personalize a language model in just a few hours using consumer hardware.
As our mystery developer highlights, “Part of what makes LoRA so effective is that it’s stackable, much like other forms of fine-tuning. Improvements, such as instruction tuning, can be applied and leveraged as contributors add on dialogue, reasoning, or tool use. While the individual fine tunings are low rank, their sum can accumulate over time, allowing full-rank updates to the model. This means that the model can be kept up to date at a low cost as new and better datasets and tasks become available, without requiring a full run.”
Consequently, generative AI is now within reach for any AI-savvy open source developer. Moreover, the open source community has excelled in utilizing high-quality, curated datasets for training, subscribing to the notion that data quality scales better than data size. These datasets are often developed using synthetic methods and sourced from other projects, maximizing their efficacy.
The recent progress made by the open source community has prompted both Google and OpenAI to reevaluate their strategies. The rapid pace of innovation combined with the absence of usage restrictions has made open source AI models an appealing alternative for many users.
This shift seems only fitting, as the success of FAANG (Facebook, Amazon, Apple, Netflix, Google) companies in the realm of Generative AI has been built upon open source AI programs. TensorFlow, PyTorch, and Hugging Face’s Transformer have been instrumental in enabling groundbreaking advancements like ChatGPT and Bard.
Interestingly, Meta, the catalyst behind this revolution, is uniquely positioned to capitalize on incorporating its code into its own products. Perhaps other leading companies, whose future hinges on AI, will recognize the advantages of allowing open source developers to work with their data models. After all, this collaborative approach has propelled virtually every major software advance of the past two decades. So why should generative AI be any different?
As our mysterious Google developer aptly states, “Directly competing with open source is a losing proposition… We should not expect to be able to catch up. The modern internet runs on open source for a reason. Open source has some significant advantages that we cannot replicate.” These words ring true, indeed.
And so, the stage is set for the open source community to disrupt the AI landscape, potentially overshadowing the once dominant tech giants. With their rapid innovation, collaborative spirit, and access to cutting-edge technology, open source developers may very well claim victory in the race for generative AI supremacy. Only time will reveal the full extent of their triumph.
Conclusion:
The rise of open source in the generative AI market presents a significant shift that could disrupt the dominance of tech giants. The ability to personalize language models on consumer hardware and the rapid pace of innovation within the open source community pose a real threat. Companies should consider leveraging open source AI models and collaborating with developers to stay competitive in this evolving landscape. Open source’s advantages and track record of success in major software advancements make it a force to be reckoned with.
