It is interesting to see how Microsoft is re-angling itself as a leader in the new generative A.I. push.
Meta has released its latest Llama 2 large language model, also known as L.L.M.
It is still in testing and has outperformed all other open-source chat models, including ChatGPT, on most benchmarks, along with helpfulness and safety.
According to Meta, Llama 2 will be made commercially available and free of charge. It will provide an alternative to the current L.L.M.s available via Google and OpenAI.
It could position Meta as a leader in the emerging A.I. development space.
As part of the new release, Meta is sharing three different versions of the model. One that has been trained on 7 billion parameters, one on 13b, and finally, a 70b version.
While releasing ‘Llama 2 Chat,’ a fine-tuned variation has got created specifically for conversational use cases.
In itself, this is a technical feat. But even more interesting is Meta and Microsoft, which have also announced an expansion of their partnership.
It will enable developers using Microsoft tools to choose between Meta’s Llama and OpenAI’s G.P.T. models while building their A.I. experiences.
Meta and Microsoft announced support for the Llama 2 family of large language models or L.L.M.s on Azure and Windows at Microsoft Inspire.
Llama 2 is designed to allow developers and organizations to build generative AI-powered tools and experiences. Both Meta and Microsoft are committed to democratizing A.I. and its benefits, and we are excited that Meta is taking an open approach with Llama 2.
Moreover, Microsoft has also invested $10 billion into OpenAI and has already built G.P.T. into most of its tools and platforms. And now, it will be plugging Llama 2 into various applications too.
Microsoft will become a key platform for connecting consumers and these leading L.L.M.s.
A key focus of Meta’s Llama 2 model is safety and making sure that the results produced by the system are accurate and limit misuse.
Considering the various issues reported with some early L.L.M.s, it could be a significant step. It includes G.P.T., which has often led users astray due to ‘hallucinations’ and sharing of misinformation or harmful perspectives.
Meta has added a significant training load around various elements to mitigate this. It includes ‘truthfulness,’ ‘toxicity,’ and’ bias.’
Based on this additional work, Meta says that Llama 2 Chat is showing great improvement over the pre-trained Llama 2 in terms of truthfulness and toxicity.
The percentage of toxic generations shrinks to effectively 0% for Llama 2-Chat of all sizes. It is the lowest toxicity level among all the compared models.
In general, when compared to Falcon and M.P.T., the fine-tuned Llama 2-Chat shows that the best performance comes out in terms of toxicity and truthfulness.
It could also make this an even more useful generative A.I. tool and rely on a broader range of tasks. G.P.T. is amazing in its capacity to produce human-like text generations.
There are also significant risks in using those outputs without checking and re-checking any references and language. To ensure that its various inputs are not negatively influencing it.
Suppose an L.L.M. could be more trusted in this respect. It could significantly expand its use case that Llama 2 is more equipped to address.
It is an interesting new consideration either way, and the integration with Microsoft will see Meta’s new L.L.M. play a bigger role in broader A.I. development. You could also see that Meta’s system eventually became a key leader in the space.
Microsoft Azure AI customers will test Llama 2 with their sample data to test and learn about its performance in different contexts.
You can read more about the Llama 2 process and dataset here.
Click here for more updates on Meta and other social media platforms.