The next big thing in the AI heat was unveiled by Meta on the 18^th of last month– the instantly revolutionary Llama 3!

‍

Four Things Make Llama 3 Immediately Worthy of Talking About:

1. It is a Large Language Model with a capacity very close to OpenAI’s proprietary GPT-4. It is said to be trained on 15 trillion tokens of training data, and is already said to have outperformed Google’s Gemini 1.5 Pro.

2. This means Llama 3 can be commercially used in the most demanding of apps and services by developers and companies worldwide.

3. And it is free of charge! Anyone can download, run, and build on it!! [though not fully ‘open’]

4. In-app usage: the Meta AI can be used in feeds, chats, and search inside Meta apps like Facebook, Instagram, WhatsApp, Messenger and even the web.

‍

1. Heightened Power!

The previous 2.0 version of Llama ( short for Large Language Model Meta AI) itself was head-turning when it was released, but when it comes to 3.0 the platform has become much more powerful. This is by virtue of the large volumes of data used for training the model and also by incorporating various new techniques for better precision and selecting a better mixture of datasets to use. Currently, Meta has released two versions, one at the 8B and the other at 70B parameter scale.

Even though Meta hasn’t released details, its launch announcement divulged that the LLM was trained on 15 trillion tokens of publicly available data – a seven times increase from Llama 2! And though OpenAI hasn’t officially revealed the volume of tokens used in training its proprietary GPT-4, it is widely estimated to be around 13 trillion. And if that can be seen as true, Llama 3 is finally in the range of the dominant GPT series. (GPT-5 is expected this summer, so…)

Turning to MMLU benchmark scores, which is a measurement of a text model's multitask accuracy, also proved promising for Llama 3. The 70B version scored 79.5 defeating other small and midsize models like Google Gemini and Mistral 7B.

‍

2. Adopting Commercially!

Along with launch announcements, Meta also shared benchmark scores of both its pre-trained and instruct models of Llama 3.

According to data shared by Meta, the pre-trained version of 70B could outsmart Google's Gemini 1.0 Pro (MMLU - 79.5 against 71.8and the Instruct model crossed the Gemini 1.5 Pro model in MMLU. It also surpassed BIG-Bench Hard (81.3 vs 75.0), and DROP, as well as HumanEval, and GSM-8K in benchmark scores.

‍

Courtesy: Meta

Engineers who are currently in the works with various products and services embedding GPT 4 models are already on the path to evaluating if Llama can practically replace OpenAI’s model. Initial download data show that developers have embraced Llama 3. The 8B Llama 3 got over 275,000 downloads in its first five days topping the current list of trending AI models on Hugging Face.

‍

Courtesy: Meta

‍

3. It's Free - Though Not Open-Source!!

2023 certainly proved to be the year of AI – OpenAI attracted millions of developers and investors alike, into the world of AI and its potential. Then came the paid version for commercial capacity ones like GPT-4 and Google’s Gemini.

But if models with commercial capacity, such as the Llama 3, which allows commercial use in most circumstances prove competitive enough, entrepreneurs and engineers may turn to the free versions – instead of paying for OpenAI or Google models. Not only Llama 3, but any new open-source models that can prove to be infinitely powerful will take the baton away from closed ones like OpenAI and Google.

Jerome Pesenti, previous vice president of artificial intelligence at Meta and now founder of a startup, Sizzle, says that embedding Llama 3 on a cloud platform like Fireworks.ai costs just one-twentieth of that accessing GPT-4 through an API.

Llama 3 also showcases the potential for making AI models smaller, so they can be run on less powerful hardware. The 8B model of Llama 3 is said to be compact enough to run on a laptop!

‍

4. In-App Usage

With the launch of Llama 3, the biggest groundbreaker it has achieved is how the Meta AI can be used in feeds, chats, and search inside the Meta apps like Facebook, Instagram, WhatsApp, Messenger, and even the web without having to leave the app in use.

You can access real-time information from across the web while using apps.

For example, if you are chatting in Messenger, you can use Meta AI in the search bar in your Messenger group chat itself to find info and data – all without leaving the Messenger app.

Similarly, one can access Meta AI when scrolling through Facebook Feed as well! You can ask Meta AI for info on anything you stumble upon in the feed right then and there – without leaving the feed and also right from the post.

‍

The 400B!!!

The tech giant goes on further in their launch announcement about plans for Llama. The company claims to release various models in the coming months that can flaunt several new capabilities like multimodality, the ability to converse in multiple languages, etc.

Similarly, a new version of Llama 3, claimed to be the largest of their models is slated for release later this year - with 400 billion parameters!

If released, the 400B version could become the first LLM available “openly” that can match the quality of its “closed” counterparts like GPT-4, Claude 3 Opus, and Gemini Ultra.

The company went on to share test result scores of the 400B version as well in their announcement – in it, the version is seen to have scored an MMLU of 86.1 – only a sliver’s difference with GPT-4’s 86.4!!

‍

Courtesy: Meta

But a new entrant is expected this summer – GPT5. Will it rob Meta of everything it has built till now in a single stroke? Maybe let’s use Llama 3 70B to prepare a prediction list?

Outside of the US, currently, Llama 3 has been released in Australia, Canada, Ghana, Jamaica, Malawi, New Zealand, Nigeria, Pakistan, Singapore, South Africa, Uganda, Zambia and Zimbabwe.

Ideas That MoveIndustries Forward

Meta’s Llama 3: What It Means for Business in the AI Race

Meta’s Llama 3: What It Means for Business in the AI Race

Ideas That Move
Industries Forward