Meta’s LLaMA doesn’t seem to have any intention of creating another GPT. As Meta explains, LLaMA is a smaller, more powerful model that offers comparable performance to competing models, but consumes less computing power and has a lower environmental impact. Moreover, unlike OpenAI, which is a proprietary licensing method, LLaMA was created in open source from the very beginning, which is also the most important feature of LLaMA.
There are four versions of LLaMA, each with 7 billion, 13 billion, 33 billion, and 65 billion parameters. The number of parameters is a factor that determines the quality of the AI response. At this time, Meta said it plans to make the model available only to research institutions under a non-commercial license, and aims to make the model available to laboratories, universities, and government agencies.
As The Guardian reports, Meta is removing one of the biggest obstacles to AI research: the enormous cost of training large-scale language models by making those models open to academia. Although OpenAI has never quantified the cost of GPT-3 training, internal estimates put each training cycle at around $10 million. Thus, universities and research institutes can save at least millions of dollars by using LLaMA.
“Access to models is granted on a case-by-case basis to maintain integrity and prevent abuse,” Meta announced. However, this meta-politics did not last long. Indeed, last weekend the entire model was made public for anyone to upload to the anonymous messaging service 4chan by a stranger.
It’s unclear what impact LLaMA’s leak will have. Indeed, the model in its current form makes almost no sense for users who lack the necessary expertise and do not have access to a high-performance computer. It’s also unclear if Meta will try to recover the leaked model in any way. However, the Guardian predicts that LLaMA will likely become the most widely used generative AI in the world.
Unlike OpenAI’s ChatGPT, free AI models like LLaMA and the open-source text image generator Stable Diffusion will compete directly with commercial services from Microsoft and Google.
Commercial service providers have invested billions of dollars in training and improving models, and want to monetize their role as gatekeepers. Companies will have to pay significant costs to create AI solutions based on these ChatGPT. On the other hand, open AI models can theoretically be built by anyone. Developers who provide services based on open AI models can also make a profit. However, the economic benefits are distributed among several actors and there is no controlling body such as a gatekeeper.
The advantage of a control-based business services approach is that accountability is clear. OpenAI, Microsoft, and Google not only generate revenue, but also ensure that their tools are used responsibly. These providers have made it clear that they will take all necessary steps to block the activities of spammers and hackers and ensure that requests and responses are filtered. However, users of Stable Diffusion or LLaMA are unlikely to adhere to these principles, so offensive or manipulated AI content could soon flood the web.
editor@itworld.co.kr
