Generative AI is evolving at a fast tempo.
Llama 2 is a generative AI mannequin, however there are a number of key variations between Meta’s Llama 2 and its rivals, together with OpenAI’s GPT-4 (ChatGPT) and Google’s PaLM 2 (Bard). Take a look at the variations right here.
2023 actually guarantees to be the 12 months of AI, with firms trying to launch their very own generative AI fashions. Meta has adopted go well with by releasing an up to date model of their LLaMa LLM mannequin known as Llama 2. Meta has launched the mannequin as open supply, that means it’s out there without cost for each business and analysis functions.
Meta and Microsoft have teamed as much as distribute Llama 2, which is now out there within the Azure AI Mannequin Catalog. This may permit builders utilizing Microsoft Azure to make use of and profit from Llama 2.
Now whereas Llama 2 is a generative AI mannequin, or extra particularly a big language mannequin (LLM), there are a number of key variations between Meta’s Llama 2 and its rivals, together with OpenAI’s GPT-4 (ChatGPT) and Google’s PaLM 2 (Bard) .
Key variations between Meta Llama 2, GPT-4 and PaLM 2:
Llama 2 much less highly effective than GPT-4 and PaLM 2 acknowledges Meta
Percy Liang, director of the Stanford Heart for Base Mannequin Analysis, notes that the Meta analysis paper acknowledges that Llama 2 is much less highly effective than GPT-4 and that there’s a efficiency hole between the 2 fashions. In benchmarks, the Llama 2 mannequin can be barely inferior to its rivals – GPT-4 and PaLM 2.
Coaching with fewer “tokens” in comparison with the competitors
For these unfamiliar with the time period, “tokens” on this context refers back to the textual content on which a generative AI mannequin was skilled. For instance, “bat” and “cool” are tokens. The extra tokens an AI mannequin is skilled on, the higher it is going to be. In line with TechCrunch, Llama 2 was skilled on two million tokens, which is double the 1.4 trillion tokens that the unique Llama was skilled on. As compared, Google PaLM 2, its direct competitor, was skilled on 3.6 million tokens, based on a Could CNBC report.
Language assist
Meta’s Llama 2 helps fewer languages than Google’s PaLM 2 and OpenAI GPT-4, with 20 languages in comparison with PaLM2’s 100 and GPT-4’s 26. Right here Google appears to be forward. Google Bard, which used the PaLM 2 mannequin, now even helps 9 Indian languages: Hindi, Tamil, Telugu, Bengali, Kannada, Malayalam, Marathi, Gujarati and Urdu. General, Google now helps over 40 languages for Bard, together with Indonesian Indonesian.
