As the battle for AI dominance continues to heat up, industry heavyweights are scrambling to formulate their own responses to OpenAI’s ChatGPT, the most rapidly adopted web application of all time.
Meta, the parent company of Facebook, just publicly released its latest foundational large language model, LLaMA (Large Language Model Meta AI), as part of its commitment to open science. The model is designed to enable researchers to advance their work in the subfield of AI and democratise access to the technology and is poised to become a key competitor to OpenAI’s ChatGPT.
LLaMA is a smaller, more performant model, making it easier for researchers who don’t have access to large amounts of infrastructure to study these types of AI models. As a foundational model, it has been trained on a large set of unlabeled data and can be fine-tuned for a variety of tasks, making it versatile for many different use cases.
As noted by Meta’s official press release, a major motivator behind LLaMA is the restricted access researchers currently have to these large language models, “hindering progress on efforts to improve their robustness and mitigate known issues, such as bias, toxicity, and the potential for generating misinformation.”
- Advertisement -
Large language models are natural language processing systems with billions of parameters that have shown new capabilities to generate creative text, solve mathematical theorems, predict protein structures, and answer reading comprehension questions, among other things.
However, full research access to them remains limited because of the resources required to train and run such large models.
The model comes in four sizes, ranging from 7 billion to 65 billion parameters (GPT-3 has 175 billion parameters), and is accompanied by a LLaMA model card that details how the model was built in keeping with Meta’s approach to responsible AI practices.
Smaller models, like LLaMA, are easier to retrain and fine-tune for specific use cases. The model was trained on 1.4 trillion tokens, and Meta chose text from the 20 most common languages, focusing on those with Latin and Cyrillic alphabets.
While LLaMA shares the same challenges as other large language models, such as bias, toxic comments, and hallucinations, it is designed to be versatile and can be applied to many different use cases. By sharing the code for LLaMA, other researchers can more easily test new approaches to limiting or eliminating these problems in large language models.
To prevent misuse, Meta is releasing the model under a non-commercial license focused on research use cases. Access to the model will be granted on a case-by-case basis to academic researchers, and those affiliated with organisations in government, civil society, academia, and industry research laboratories around the world.
Meta believes that the entire AI community must work together to develop clear guidelines around responsible AI, and responsible large language models, in particular. With the release of LLaMA, the company hopes to enable researchers to better understand large language models and to encourage further research in this crucial area.
Bias and Toxicity
AI is rapidly transforming many industries, from healthcare to finance to transportation. However, as AI systems become more prevalent in our lives, developers must be aware of the potential for bias and toxicity to impact the way these systems function and the outcomes they produce.
- Advertisement -
Bias in AI can be caused by a variety of factors, including the data used to train the system, the algorithms used to process that data, and the way the system is deployed. For example, an AI system used to determine creditworthiness may be biased against certain ethnic or racial groups if the data used to train the system is not diverse and representative. This can lead to unfair practices and discrimination against these groups.
Toxicity in AI, on the other hand, can manifest in many ways. It can include offensive language, hate speech, or discriminatory content generated by AI systems. For example, a chatbot designed to interact with users online may generate offensive language or content if it is not properly programmed to recognise and avoid such behaviour.
To address these challenges, it is crucial to create diverse and representative datasets that accurately reflect the populations the AI systems will interact with. It is also important to ensure that ethical guidelines are followed in the design and deployment of these systems, including addressing issues of bias and toxicity.
In addition, ongoing monitoring and evaluation of AI systems can help identify and address any bias or toxicity that emerges over time.
This includes continually updating and refining the algorithms used to process data, as well as implementing safeguards to prevent the generation of toxic or offensive content.
Ultimately, the successful development and deployment of AI systems will require ongoing attention and effort to ensure that these systems are fair, ethical, and responsible. By addressing bias and toxicity in AI, we can build systems that benefit everyone and help create a more equitable and just society.