Innovation with Open Weights
The release of Meta's newest open-weight LLM model, Llama 3.1, could be a source of concern for OpenAI and several other companies with closed-source models. For the first time, an open-source model is among the top performers in the standardized evaluation rankings. For context, all closed-source models (Anthropic Claude, OpenAI, Cohere, etc.) are consumed via APIs, where the vendors maintain the infrastructure and sell usage fees (in this case, the unit of measurement is Token).
Once we have a model of equal or superior quality that can be freely used running on our own hardware, things change completely. Each company will be able to develop its own generative AI innovations in-house, dramatically reducing the cost of inference. This poses a serious threat to OpenAI, as most of what we see in the AI application market today are nothing more than companies consuming OpenAI's API (ChatGPT) and embedding them in applications used for a wide variety of purposes. With a model as good as Llama 3.1, which can be run in-house, these companies implementing these applications will have their models at cost price.
The Power of the Llama 3.1
Llama 3.1 represents a significant breakthrough in language modeling. Developed by Meta, this model not only supports multilingualism but also features advances in coding, reasoning, and tooling. With 405 billion parameters and a context window of up to 128 tokens, Llama 3.1 is on par with market leaders like GPT-4.
The robust infrastructure used to train Llama 3.1, which includes clusters of H100 GPUs, ensures the efficiency and stability required to develop a model of this magnitude. Furthermore, Llama 3.1 is being expanded to include multimodal capabilities such as image, video, and speech recognition, making it an even more powerful and versatile tool for a variety of applications.
Challenges for OpenAI
On the other hand, OpenAI faces significant financial challenges. According to an analysis by The Information, OpenAI could lose up to $5 billion this year due to high AI training and inference costs, which could reach $7 billion. Furthermore, personnel costs could reach $1,5 billion. This critical situation puts OpenAI in a vulnerable position, especially with the need to raise more capital in the next 12 months.
The Race for Hardware
Now, more than ever, the race will be on to find the hardware to support all of this. While Nvidia is by far the champion in this market, AMD, AWS, Intel, and several others are also developing their own chips for training and inference. A company's ability to develop and maintain its own hardware infrastructure will be a crucial factor in its ability to compete in the AI market.
Paradigm change
The availability of an open model like Llama 3.1, which can run internally on company infrastructures, represents a paradigm shift. Companies will be able to drastically reduce their inference costs while simultaneously increasing their capacity for innovation in generative AI. This will boost the market for generative AI applications. IA, which currently relies heavily on OpenAI's APIs, can begin to migrate to more cost-effective and efficient in-house solutions.
Conclusion
In short, Llama 3.1 not only promises to revolutionize the field of language models, but also puts significant pressure on OpenAI and other closed-source model providers, changing the AI landscape in profound and lasting ways. With the hardware race intensifying, the future of AI will be shaped by both software innovation and hardware supportability, bringing new challenges and opportunities to the industry.
