In a major escalation of the AI competition, Meta has unleashed two new entries in its Llama series of generative AI models: Llama 3 8B and Llama 3 70B. The company boldly claims these models represent a significant leap over previous versions and even outperform leading models from Google and others.
Meta’s claims hinge on the number of parameters in each model (essentially a measure of AI skill) and performance on popular, though somewhat controversial, AI benchmarks. Llama 3 8B and Llama 3 70B are said to be among the best-performing models available, with Meta claiming its larger Llama 3 70B rivals Google’s flagship Gemini 1.5 Pro and outperforms Anthropic’s lesser-known Claude 3 Sonnet model.
To support these claims, Meta touts a massive training dataset of 15 trillion tokens (words and code) and a self-developed testing suite. However, they remain secretive about the data’s origin – raising concerns about potential copyright infringement, a known issue in generative AI. Meta’s internal test results should also be taken with skepticism.
Meta promises that the new Llama models are more accurate, less toxic, and less likely to refuse questions. However, they’ve made similar claims in the past, with mixed results. Rigorous testing by independent researchers will be crucial to determine if the Llama 3 models truly address the shortcomings of previous iterations.
While Meta markets the Llama 3 models as “open,” their release terms are far from it. Developers can use the models for research and commercial purposes but are barred from building competing AI models with them. Those with large user bases must also seek special permission from Meta, limiting accessibility and raising questions about Meta’s commitment to true openness in AI.
