Meta is releasing LLaMA to compete with PaLM and Chinchilla. Will it suffer the identical fate as Blenderbot and Galactica? Or the third time is a charm?
You might have heard all the thrill these days about language models developed by big names like Microsoft, Google, and OpenAI. But let’s not ignore Meta, Facebook’s parent company, which remains to be attempting to make an announcement. Today (well, yesterday), they released a brand-new AI language generator called LLaMA. Well, as recent as it could actually get!
LLaMA shouldn’t be like ChatGPT, no it shouldn’t be a Chatbot, but a successor of Galactica. As an alternative, it’s a tool for researchers to assist discover and tackle the issues related to AI language models. Meta claims that this model will help researchers to tackle issues like bias, toxicity, and the tendency to fabricate information.
By releasing LLaMA, Meta hopes to “democratize access” (Cliche much?!) to this critical field of study. They’re making the tool available under a “noncommercial license focused on research use cases,” meaning universities, NGOs, and industry labs can access it. Meta believes that your complete AI community, from academics to policymakers (ahem!), must work together to develop clear guidelines around responsible AI.
LLaMA research paper by Meta claims that their second-smallest version, LLaMA-13B, beats OpenAI’s GPT-3 model in most benchmarks, and their largest brother of this model, LLaMA-65B, is just nearly as good as the perfect models on the market, like DeepMind’s Chinchilla70B and Google’s PaLM 540B.
Now, here’s the cool part — once LLaMA-13B is trained, it could actually run on only one Nvidia Tesla V100 GPU. So if you’ve $8000 (amazon) to shell out, you need to use it at your private home too! But in all seriousness, that’s great news for smaller organizations who wish to check it out, but not a lot for individual researchers who may not have access to that sort of apparatus. I will likely be covering the small print of the research paper in a separate article but until then, check these benchmarks out
Interestingly, Meta’s release didn’t get as much attention as another AI chatbots on the market. I can understand, though, given how badly their Galactica screwed up.! Which may actually be an excellent thing.
Meta has actually released a few of their very own chatbots previously, but they didn’t accomplish that hot. One, called BlenderBot, was just about just not superb, and one other one, Galactica, was presupposed to write scientific papers but ended up producing a bunch of nonsense. Yikes!
Based on Mark Zuckerberg, Meta’s CEO, LLaMA is a state-of-the-art model designed to assist researchers with their work. This baby can generate text, have conversations, summarize written material, and even tackle more complex tasks like solving math theorems or predicting protein structures. Pretty impressive, huh? Seeing the way it pans out in the true world will likely be great.
Meta is hoping for a greater reception with the discharge of LLaMA — and who can blame them? The third time is a charm 😉 With a model this advanced and groundbreaking, it’s hard to not get excited. Here’s to the long run of AI and all of the incredible advancements it would bring!
