Meta and Microsoft launch Llama 2 artificial intelligence model: Here are all the details

Meta announced Llama 2, a new family of artificial intelligence models designed to drive applications like other modern chatbots, especially ChatGPT developed by OpenAI. Trained on a mix of publicly available data, Meta claims that the performance of the Llama 2 has improved significantly over previous generation Llama models. Llama 2’s trademark seems to be safer than other models.

Llama 2 is commercially available

Llama 2 is a successor to Llama, a collection of models that can generate text and code in response to commands, comparable to other chatbot-like systems. However, not everyone who wanted could use the Llama model. Meta had decided to block access to the models for fear of abuse, but the model was later somehow leaked onto the internet. In contrast, Llama 2 will be free for research and commercial use.

Meta says the Llama 2 is optimized for Windows as well as smartphones and PCs that host Qualcomm’s Snapdragon chips and will be easier to operate, thanks to an expanded partnership with Microsoft. (Qualcomm says it’s working to bring the Llama 2 to Snapdragon devices in 2024).

How is Llama 2 different from Llama?

Llama 2 has two variants, Llama 2 and Llama 2-Chat; the latter is finely tuned and optimized for two-way conversations. Llama 2 and Llama 2-Chat come in versions of different complexity: 7 billion parameters, 13 billion parameters, and 70 billion parameters.

Llama 2 is also trained on two trillion tokens, where “tokens” represent raw text. That number is nearly double the number the Llama was trained at (1.4 trillion), and generally speaking, the more tokens the better when it comes to AI. Meanwhile, Google’s current flagship major language model (LLM) PaLM 2 is reportedly trained on 3.6 trillion tokens, and OpenAI’s GPT-4 model is estimated to be trained on trillions of tokens.

Meta does not disclose specific sources of Llama 2’s training data. The company states that these were collected from the web, mostly in English, but not from the company’s own products or services, and they highlight “factual” text. Frankly, we can say that the main reason why “educational data” is not disclosed when it comes to artificial intelligence is legal discussions, because it seems like there will be a big copyright struggle.

Returning to the topic, Meta says that the Llama 2 models perform slightly worse than its biggest rivals, the GPT-4 and PaLM 2, and the Llama 2 lags significantly behind the GPT-4 in computer programming. However, Meta claims that human evaluators have found Llama 2 roughly as “useful” as ChatGPT

Meta also acknowledges that Llama 2, like all generative AI models, has biases in certain axes. Llama 2, for example, tends to produce “male” pronouns at a higher rate than “female” pronouns due to imbalances in the training data. As a result of the toxic texts in the training data, it does not outperform other models in toxicity benchmarks.

Finally, due to the nature of open source models, it is not possible to say exactly how or where the models will be used. Considering that the internet moves at lightning speed, it won’t be long before we find out.