LLaMA 2: Meta presents new AI language model
The US technology company Meta has presented a new version of its AI language model. LLaMA 2 is open source and has roughly the same performance as GPT-3.5.
Together with Microsoft, Meta has presented LLaMA 2. The second version of the generative language model is open source; according to "Financial Times", Meta is planning a commercial version with a paid licence. This means that LLaMA 2 could soon find its way into software products. The collaboration with Microsoft relates to distribution via Azure on the one hand and optimisation for local operation of LLaMA 2 under Windows on the other.
According to Meta's own information, the artificial intelligence (AI) is significantly more powerful than in the first version. In benchmarks, it is roughly on a par with OpenAI's GPT-3.5, but far behind the new GPT-4. Further details can be found in Meta's research report.
Source: Instagram / Mark Zuckerberg
LLaMA 2 is available in three different versions: with seven billion, 13 billion or 70 billion training parameters. It is also intended to be particularly suitable as a programming tool. The language model will "change the landscape of the LLM market", writes Meta's Head of AI Yann LeCun on Twitter. It is available on Microsoft Azure and Amazon Web Services (AWS), among others. Access can be requested via this form.
With its open source model, Meta is swimming against the tide of other major AI suppliers such as OpenAI and Google. The advantage of this is that companies can further develop LLaMA 2 themselves. However, this raises security concerns. The AI could be misused for dubious purposes, such as disinformation or spam.
What Meta does not reveal is where the training data for LLaMA 2 comes from. This is also a source of criticism, as such data often comes from the internet without authorisation. In Meta's case, it could also come from the social networks Facebook and Instagram's own databases.
Own image generator
On Monday, Meta had already presented an AI product: CM3Leon, a generative AI model for image generation and image description. It can therefore do text-to-image and image-to-text. For example, an image can be fed into the AI and then modified using a text command.
Source: Meta
12 people like this article
My fingerprint often changes so drastically that my MacBook doesn't recognise it anymore. The reason? If I'm not clinging to a monitor or camera, I'm probably clinging to a rockface by the tips of my fingers.