Hi, this is Wayne again with a topic “Meta claims Llama 3 is the most advanced open source AI yet l TechCrunch Minute”.
How many llamas will meta need to win the AI race? Well, this week, meta added two more models into its generative: AI family, the llama 38b and llama 370b. Models are joining llamas 1 and two. This massive training set should make them smarter, quicker and allar around better for the user, at least that’s what meta is saying. The new llama models were trained on two enormous clusters of gpus. In fact, meta says that for their underlying data, the llama 38b and llama 370b models are among the best performing generative AI models available today.
But of course meta is going to say that the good news is that there are objective measures out there to test the company’s claims and across at least nine benchmarks. Llama 38b bested other open- source models, including one from French company Mistral the mistol 7B model or Google’s Jimma 7B model. These latter two models aren’t the most Cutting Edge out there, but the tests do show that meta’s AI Champion is absolutely no slouch. The company is going to cram AI into every nook and cranny of its Facebook, Cash Cow, for example, AI will be increasingly tied to search and chat, and the company is going to give users even more powerful tools to create images from prompts.
In addition, meta says that users of the new llama models should expect a lower likelihood of them, refusing to answer questions and higher accuracy on trivia questions and questions pertaining to history. Amongst other things, if you thought your feed was fillmed with AI, now just wait. It’S going to get worse. Why are we talking about two new models in the AI world when it seems that new models come out every 27 minutes? Well, there’s a Twist to what meta has built.
The Llama AI models are open source, which means that they’re available for free and can be used and Modified by anyone. While this is an amazing way to build quickly, there are some out there who think that the open source AI approach is dangerous. They argue that open- Source models leave the door well open for Bad actors to access and potentially misuse powerful AI tools.
On top of all that, training data is seen by AI companies as a competitive Advantage. So it tends to be a little bit of a black box, so while the Llama models are open, source meta won’t say where the training data for its new models came from. It just says that they were built using quote publicly available sources and given how many articles I’ve written on the internet over time, I can’t help but wonder if any of my words were hoovered up by meta to help build build these latest llama models.
I guess the answer is probably one more thing before we go. These models must also play nice. Medic claims that the new llama models have reduced toxicity bias due to a team of filters that it put into place, but filters are absolutely not foolproof, so we’ll have to wait and see how the Llama 3 models act when they’re actually out in the wild AI.
It never stops I’ll see you tomorrow. .