As early as June, we'll see if AI becomes too bold due to the new approach to security.
Meta is preparing to release the next version of its AI model called Llama 3, which, according to the company, should become more responsive to the user and the context of communication.
Unlike the conservative approach in Llama 2, where Meta avoided any PR scandals, the new Llama 3 is designed to subtly distinguish between the dangerous and harmless meaning of words with a double meaning. To do this, the team has a specialist who will oversee the training of the model and be responsible for making its responses more flexible. However, just over the past month, three leading experts in the field of AI security left Meta* at once, so it is not known how this will affect the quality of development.
It is not reported whether Llama 3 will become a purely language model, like Llama 2, or a multimodal one that can understand and generate not only text, but also images. According to available data, this model, which is being developed as an alternative to GPT-4, can contain up to 140 billion parameters. This is about twice as much as in the previous version. However, 140 billion parameters is only a fraction of the declared volume of GPT-4, which was originally supposed to be even larger.
Training for Llama 3 has not yet begun, but it is expected that it will still reach the GPT-4 level in terms of generation quality. Recall that the release of Llama 2 took place in July last year.
In general, despite the outflow of staff, Meta continues to adhere to ambitious plans in the field of generative artificial intelligence. It intends to maintain a relatively open development strategy. The head of Meta, Mark Zuckerberg, is one of Nvidia's biggest clients. By the end of this year, he plans to use about 600 thousand Nvidia video cards for training neural networks.
The ultimate goal of Meta is to create general-purpose artificial intelligence that can solve a wide range of tasks. This is consistent with the principles of OpenAI.
Currently, there are two main approaches to ensuring the security of artificial intelligence.
The first is highly regulated, which is used by companies such as OpenAI in GGPT and GPT-4, as well as Google in the Google Advanced model. Here, models are given clear rules for polite and politically correct responses.
The second is a more free approach, implemented in open source projects and in Elon Musk's Grok program. Unfortunately, Grok is not yet reliable enough in terms of predictability and security for widespread use.
The original GPT-4 Mix-of-Experts model, with which Llama 3 is compared, had 1.76 trillion parameters. Although parameters are no longer the only and most important indicator of the quality of neural networks.
Meta is preparing to release the next version of its AI model called Llama 3, which, according to the company, should become more responsive to the user and the context of communication.
Unlike the conservative approach in Llama 2, where Meta avoided any PR scandals, the new Llama 3 is designed to subtly distinguish between the dangerous and harmless meaning of words with a double meaning. To do this, the team has a specialist who will oversee the training of the model and be responsible for making its responses more flexible. However, just over the past month, three leading experts in the field of AI security left Meta* at once, so it is not known how this will affect the quality of development.
It is not reported whether Llama 3 will become a purely language model, like Llama 2, or a multimodal one that can understand and generate not only text, but also images. According to available data, this model, which is being developed as an alternative to GPT-4, can contain up to 140 billion parameters. This is about twice as much as in the previous version. However, 140 billion parameters is only a fraction of the declared volume of GPT-4, which was originally supposed to be even larger.
Training for Llama 3 has not yet begun, but it is expected that it will still reach the GPT-4 level in terms of generation quality. Recall that the release of Llama 2 took place in July last year.
In general, despite the outflow of staff, Meta continues to adhere to ambitious plans in the field of generative artificial intelligence. It intends to maintain a relatively open development strategy. The head of Meta, Mark Zuckerberg, is one of Nvidia's biggest clients. By the end of this year, he plans to use about 600 thousand Nvidia video cards for training neural networks.
The ultimate goal of Meta is to create general-purpose artificial intelligence that can solve a wide range of tasks. This is consistent with the principles of OpenAI.
Currently, there are two main approaches to ensuring the security of artificial intelligence.
The first is highly regulated, which is used by companies such as OpenAI in GGPT and GPT-4, as well as Google in the Google Advanced model. Here, models are given clear rules for polite and politically correct responses.
The second is a more free approach, implemented in open source projects and in Elon Musk's Grok program. Unfortunately, Grok is not yet reliable enough in terms of predictability and security for widespread use.
The original GPT-4 Mix-of-Experts model, with which Llama 3 is compared, had 1.76 trillion parameters. Although parameters are no longer the only and most important indicator of the quality of neural networks.
