Meta, Facebook’s parent company, recently launched Llama 2, a free and open-source artificial language model that promises to revolutionize the field of artificial intelligence. Llama 2 is based on the source code of GPT-3, the world’s most advanced language model, developed by OpenAI, a non-profit organization backed by Elon Musk and other technology industry figures. Llama 2 is intended to give developers and companies easy and free access to cutting-edge artificial intelligence without having to pay OpenAI’s high fees.
In July, Meta made its Llama 2 model available for commercial use through partnerships with major cloud service providers, including Microsoft. Meta does not charge any fees to access or use the model it develops, the company said. Rather, by opening up other companies’ technology, Meta will benefit from the improvements that can be achieved if more developers use it, stress test it, and identify the problems associated with it.
Llama 2 is intended to give developers and companies easy and free access to cutting-edge artificial intelligence without having to pay OpenAI’s high fees.
However, some OpenAI customers who tried Llama 2 found that using the free model wasn’t as affordable. In fact, running Llama 2 requires a lot of computing resources, resulting in high cloud computing costs. According to a study by Baseten, a startup that helps developers use open source language models, using Llama 2 out of the box is 50 to 100% more expensive than using GPT-3.5 Turbo, an OpenAI model, which supports services like ChatGPT. The open source option is only cheaper for companies that want to customize a language model by training it on their data; In this case, a custom Llama 2 model costs about a quarter of the price of a custom GPT 3.5 Turbo model, Baseten noted.
An example of this cost difference is provided by Andreas Homer and Ebby Amir, co-founders of Cypher, an application that helps people create virtual versions of themselves in the form of a chatbot. These two entrepreneurs tested Llama 2 for their application, which earned them a $1,200 bill from Google Cloud, their cloud provider, in August. Then they tried GPT-3.5 Turbo and were surprised that it only cost them $5 per month to complete the same amount of work.
Baseten also noted that OpenAI’s most advanced model, GPT-4, is about 15 times more expensive than Llama 2, but is generally only needed for the most advanced generative AI tasks such as code generation, and not for those that require the Most large companies want to integrate.
These results suggest that, despite its stated desire to make AI accessible to everyone, Meta has so far failed to compete with OpenAI in the artificial intelligence market. It seems that the quality and effectiveness of OpenAI’s language models still justify their price and that companies looking to use AI for their needs must weigh the pros and cons between free and paid options.
Several model sizes are available
Llama 2 is an artificial language model that uses machine learning to generate text from a specific keyword, phrase, or context. Billions of texts from public sources such as the Internet, books, articles, etc. are analyzed and the rules and patterns of natural language are learned. He can then use this knowledge to produce original and relevant texts, imitating the style, tone and content of the original sources.
Llama 2 is based on the GPT-3 source code developed by OpenAI. It uses the same architecture, called a transformer, which consists of a series of layers of artificial neurons that process text in parallel. It differs from GPT-3 in that it is trained on a more diverse and up-to-date dataset, covering over 100 languages and different areas such as chat, coding, poetry, etc. It also benefited from task-specific fine-tuning based on human-annotated examples to improve performance on specific applications.
There are three model sizes available: 7 billion, 13 billion and 70 billion parameters. The larger the model, the more powerful it is, but the more computing resources are required to run it. Llama 2 can be used for a variety of natural language processing tasks such as summarizing, translating, question answering, and more. It can also be customized to users’ needs by training them based on their own data and tasks.
OpenAI is making several attempts to foray into the professional world
OpenAI announced in March that APIs for its ChatGPT and Whisper models are now available, giving developers access to AI-powered voice and text-to-speech capabilities. Through system-wide optimizations, OpenAI has successfully reduced the cost of ChatGPT by 90% since December and passed these savings on to API users. OpenAI launched Whisper, a text-to-speech model, as an open source API in September 2022. The Whisper API has received high praise from the developer community. However, it can be difficult to use.
OpenAI has now introduced GPTs, a way for anyone to create their own version of the popular conversational AI system. Not only can you create your own GPT for fun or productivity, but you’ll soon be able to publish it on a marketplace called the GPT Store… and maybe even make a little money in the process.
The custom GPTs were announced on Monday at DevDay, OpenAI San Francisco’s first-ever developer conference, where the company also announced a turbocharged GPT-4 and cheaper, lower pricing for developers using its models in their apps, as well as the news that ChatGPT has reached an incredible 100 million weekly users.
We believe that when you give people the tools, they will achieve incredible things,” said founder and CEO Sam Altman on stage.
To that end, the company is introducing so-called GPTs, “customized versions of ChatGPT that you can create for a specific purpose.” It is possible that the chosen name causes confusion, as GPT (Generative Pretrained Transformer) is actually the technical term for this type of large language model.
These initiatives aim to strengthen OpenAI’s position in the artificial intelligence space and build customer loyalty against competing offerings such as Meta’s.
The battle between Meta and OpenAI for control of artificial intelligence is therefore not over and could even intensify in the coming months with the introduction of new models and new applications. Users and businesses must make a clear choice between the various options available, taking into account their needs, budget and values. The future of AI will largely depend on these decisions and how they influence the development and use of this technology.
Sources: Meta, Baseten study
And you ?
What do you think of Meta’s strategy of making its Llama 2 language model free and open source? A good initiative to democratize artificial intelligence? For what ?
Are you generally for or against open source development of AI?
Some OpenAI customers who tried Llama 2 found that using the free model wasn’t as affordable. Are you surprised?
What are the advantages and disadvantages of using Llama 2 compared to OpenAI’s language models, such as GPT-3, GPT-3.5 Turbo or GPT-4? What criteria do you use to select the most suitable model for your needs?
How do you assess the quality and reliability of the texts generated by Llama 2 and competing models? Have you ever encountered any problems or errors with these templates? How did you solve them?
What apps do you use or plan to use with Llama 2 or competing models? What benefits do you get from this or do you hope to gain from it?