AI Personas will be the first mass-use tool to demonstrate the capabilities of LLaMA, Meta’s new AI-powered language model.
The Mark Zuckerberg-led company plans to enable a series of AI-powered personas within its ecosystem of products and services, which will serve as virtual assistants for users of apps such as Instagram, Facebook, or WhatsApp.
The company’s newest project will be powered by LLaMA, Meta’s new open-source AI that seeks to compete against OpenAI’s GPT-4. The novelty of an open-source AI caught the attention of developers around the world, who claim that this code will facilitate the creation of new and powerful systems.
Let’s find out all about AI Personas, LLaMA, Meta algorithms, and the future of artificial intelligence as we know them!
Personal assistants in Meta apps
At the end of February, Meta CEO and co-founder Mark Zuckerberg announced that the company was putting together an internal team focused on developing a new artificial intelligence. Zuckerberg detailed that they were developing a group of products based on generative AI to enhance their work in this area. He also communicated that they convened all the teams working on Meta’s generative AI to bring them together in a single group.
During the interview, the CEO talked about the new personal assistants that users will be able to use in Meta applications. The “AI Personas”, as they are called by the company, will have different personalities according to the tasks for which they were designed. The functions promise to help users get recommendations based on their interests, search for content, or interact with all the big tech platforms.
In the short term, Meta intends to develop creative and expressive tools based on generative AI. However, its long-term goal is to complete the development of “AI Personas” so that AI-powered personal assistants will be able to help users of Meta’s apps. According to Mark Zuckerberg, the company is exploring various features of generative AIs to offer a better texting experience on platforms such as WhatsApp or Facebook Messenger; being able to integrate intelligent tools to create images or videos that can be posted directly to Instagram or Facebook. Similarly, multimodal resources would be added for the entire ecosystem comprising Meta’s network.
In early July, app researcher and mobile developer Alessandro Paluzzi, known in the tech scene for being a reliable leaker, reported that he found a new feature in Instagram’s source code that would allow users to “chat with an AI” without leaving the app. Accordingly, the AI system would have the ability to offer advice, compose short texts o, and answer queries using 30 different personality types.
How does Meta’s AI Personas work?
In the quarterly presentation of Meta’s financial results, Mark Zuckerberg expressed that the new resource they were developing would as its main function to help users connect and express themselves better in the company’s applications. To achieve this, they will add creative tools that make sharing content more fun and easier.
“AI Personas” will act according to the needs of the users, being able to act as coaches, assistants, helpers, or instructors. In this way, it will be easier and more entertaining for users to interact with other creators or companies. Zuckerberg informed that the advances in the artificial intelligence sector and the development of “AI Personas” are based on their new language model, which they called LLaMA.
When the launch of “AI Personas” becomes a reality, the company will gain a new resource with which to compete in the artificial intelligence market. The arrival of a system of virtual assistants based on generative AI in applications such as Instagram, WhatsApp, or Facebook will strengthen the functionality portfolio of Meta’s platforms; allowing the company to compete against the popularity of TikTok. The launch of these chatbots would reflect the enormous capacity of Meta’s new AI language model; this being a relatively new system that has barely updated to its second version.
On the other hand, the British newspaper Financial Times pointed out that Meta’s algorithms could become the new source of informative data about the users of the company’s platforms. Since last year Meta started collecting information about subscriber profiles on all its platforms, but the new algorithm would include even more precise advertising tools and formats never seen before.
Meta’s new AI
Meta launched LLaMA earlier this year, an acronym that stands for Large Language Model Meta AI and translates to Large Language Model Meta AI. This new language model is made up of four artificial intelligence systems that operate under a non-commercial license and are geared toward research uses. According to its developers, the models will be accessible to groups such as NGOs, universities o, and laboratories whose projects are focused on optimizing the use of artificial intelligence
Initially, LLaMA was different from other language models already known, because it was focused on a different market sector; therefore, it did not compete directly with other systems created by companies such as Microsoft, OpenAI, or Google. However, a Financial Times report states that Meta plans to launch a second commercial version of its AI, which will be focused on the enterprise market and will offer a personalized experience. In this way, the big tech would be considering entering a more competitive niche.
The commercial version of LLaMA will be free and open source for start-ups and developers interested in creating customized versions of the model. In this way, the company seeks to provide an affordable alternative for independent programmers to build new products or services taking advantage of the AI boom. However, the company plans to charge users who wish to modify the language model to create their proprietary products. An open source for a commercial version brings with it risks and advantages, so these measures are aimed at safeguarding Meta’s investment. Since its foray into the artificial intelligence market, the big tech’s strategy has focused on open-source projects, as they will enable the company to expand its level of competitiveness.
LlaMA 2: The new commercial open-source language
Meta officially launched LLaMA 2, the enhanced version of its new artificial intelligence-based language model, which is targeted for commercial use and will come in three different sizes. The new service was announced in early 2023, as a set of four AI models exclusively available to educational institutions, researchers, and non-profit organizations.
Unlike the first version of the language model, LLaMA 2 was released from the beginning under an open-source system and free for commercial use. According to Meta, the second version has 40% more data and can create chatbots capable of generating results as good as OpenAI’s ChatGPT.
During Microsoft Inspire, an event held for partners seeking to transform the AI ecosystem, Meta announced that it would release the commercial version of its new language model. Earlier this year, the two companies partnered to extend access to LLaMA 2 through the Azure AI catalog; Microsoft’s artificial intelligence service that provides solutions to data scientists and developers. For this reason, the model is adapted to run natively on Windows OS devices and offers interested users the option to download it to their computers.
According to Mark Zuckerberg’s statements, the LLaMA 2 version promises to be more powerful than its predecessor; the new AI was trained with 40% more parameters and data that guarantee twice as much context processing. Pre-trained versions of the model were trained with more than 1 million human inputs and 2 billion tokens. Also, LLaMA 2 is available in three different sizes: 7,000, 13,000, and 70 billion parameters.
The developers have emphasized that the new model is commercially licensed open source, a feature that could revolutionize competition in the creation of new artificial intelligence-based applications. However, LLaMA 2’s “open source” label is not strict. Meta’s latest AI model can be used free of charge by third parties to develop and modify the operation of the system and can be adapted by other companies to create their tools for profit-making purposes. In the first instance, the criteria match the definition of what open-source software is supposed to be. However, the commercial use license of LLaMA 2 has a restriction that contradicts the open-source guidelines.
In the LLaMA 2 whitepaper, in the Additional Commercial Terms section, Big Tech states that if the licensee’s services or products reach more than 700 million active users per month, they must request a special license from Meta, who may do so at their discretion. In addition, the licensee may not exercise any rights under the agreement unless Meta expressly authorizes the exercise of such rights.
The risks of open-source AI
In the first month, Meta seems to have outperformed companies with more robust models and structures, at least in terms of speed and volume of release. The flexibility and ability to offer continuous improvement on each system, involving hundreds of prestigious developers, has quickly made Meta’s new open-source language model one of the biggest rivals to companies like OpenAI or Google.
For Meta, the simple fact of making inroads into the ecosystem, transforming i,t, and expanding its catalog of developments to be adopted by the market is the win. Turning investments, license requests from other companies, and agreements into collateral benefits.
The strategy proposed by Big Techcano completely changes the regular flow of the market; since the company has the necessary resources to develop dozens of language models and launch them at an early stage at a low cost. Being open-source models, external developers can access the code and make improvements to their AI tools. In this way, third parties can create new products for their use and profit; while Meta gets additional benefits or features for its ecosystem.
However, there are several risks that Meta will have to face. The degree of openness that comes with an open-source language model, coupled with the lack of worldwide regulations, provides the ideal conditions for the system to be used to create malicious or illegal content.
The regulatory risks surrounding the intellectual property of the resources used to power and refine the system’s operation can trigger a series of legal problems for the company. Recently, humorist Sarah Silverman sued Meta, Microsoft, and OpenAI alleging that the companies used her presentations without her consent to train artificial intelligence models.
As of late 2022, a series of laws are being drafted that will aim to regulate the scope, use, and development of artificial intelligence. For the time being, it is likely, that remedies such as the Disclaimer available in the commercial version of LlaMA will not be sufficient to avoid legal problems.
LLaMA 2: The quasi-open-source AI
Meta’s contradiction was pointed out by Open Source Initiative, a technology organization that celebrated through a Twitter post the launch of the new model; although it publicly asked the company to “watch its language” so as not to classify its AI as fully open source.
The organization, which coined the term “open source” and is responsible for educating on this subject on a non-profit basis, pointed out that the LLaMA 2 license only authorizes some commercial uses, as indicated in the whitepaper. However, the term “open source” has a clear definition that excludes any restrictions on commercial use.
Although the definition is familiar to everyone involved in the industry, the clarification makes Meta’s intentions clear between the lines. The big tech is aware of its ability to explore and capitalize on the open-source ecosystem, but its goal is to position itself among the first places in the community of researchers and developers who do not have the financial or technical backing to access the APIs of models such as PaLM 2 or GPT-4. Even so, the company led by Mark Zuckerberg does not plan to allow its AI model to be manipulated so freely by organizations trained to develop systems that compete against its system; as they could put its business at risk.
Google’s Concern and the LLaMA Leak
In early May, an anonymous document allegedly written by a Google researcher was leaked, expressing concern about the future of the technology giant. The text pointed out that, while executives were focused on finding measures to counter the competitive threat caused by OpenAI’s text generation technology, open-source software was “snatching” a large share of the market. The paper cited LLaMA and described it as a new language model developed by Meta, to which only a few researchers had access by invitation.
However, within days the code was leaked on the 4Chan website, quickly becoming popular with programmers who took advantage of the software to adapt and build systems from the development. Weeks after the leak, variants of LLaMA were released called Vicuna and Alpaca; which were almost as accurate as ChatGPT but with enough flexibility to be customized on a computer. The leaked Google document reflected their concern about this, stating that the impact on the artificial intelligence ecosystem was considerable and that anyone now could experiment.
Probably, the sum of all these events related to the first version of LLaMA caused the big tech to bend the rules of open source at the time of launching the second version of the language model. Consequently, the commercial license of LLaMA 2 is a Meta measure to control the large projects that are developed from its code, as well as who will keep the profits; actions with which the company intends to protect its business.
At the moment, Meta’s route to discovering and attracting new AI talent is based on: collaboration with the research and scientific community, a smaller structure,e and cooperation with independent developers to perfect their projects in the artificial intelligence ecosystem. Similarly, the company anticipates that commercial licensing of LLaMA 2 will enable it to describe new uses for its AI; which could generate revenue in the long term.
LlaMa 2: Is it an open-source AI?
Meta’s VP of Generative AI, Ahmad Al-Dahle, recalled during an interview that the company has always championed open-source developments. In fact, in 2016 the Facebook AI Research Lab launched the open-source machine learning library PyTorch; which is very popular among independent developers. One of the company’s motivations for creating a community around open source came from the fact that they noticed the high demand beyond researchers to work on and improve these models.
Now, while the LLaMA 2 language model lends credibility to Meta as a leading exponent in open-source, not all elements of the release fit the category. The data used to train the models is only described in the release documentation as “publicly available online sources,” with no further details related to the elements used during its development.
The commercial license used by Meta in LLaMA 2 requires licensees with more than 700 million monthly active users to enter into a separate implementation agreement with Meta. The reason is clear, the clause provides Meta with insurance in case other technology giants wish to build on its system. The model language also includes a Permitted Use Policy, which prohibits creating malicious code, encouraging criminal activity, promoting violence, harassment o, or abuse using LLaMA as a source. However, the company has not issued statements about what action it would take if a licensee violates any of these policies.
So, is LLaMA an open-source language model? According to the Open Source Initiative, Meta’s language model should not be qualified by this definition; because its usage policies are out of step with the principles of open source systems. Although the LLaMA code is promoted for free use and access to various types of developers, it does not meet all the conditions of an open-source system. Therefore, based on the strict definition of the term and the position of the Open Source Initiative organization, we can affirm that the new Meta language model is not open source; although it could be considered as a partially open source.
Future of the market: The implications of open source AI.
“LLaMA 2 is the first open-source language model capable of competing directly against ChatGPT,” says Nathan Lambert, an AI researcher at Hugging Face, a company that publishes open-source generative models and machine learning systems. Although Lambert does not consider the development to be “authentically open source,” he is surprised by the large number of variations of LLaMA 2 that he has found on social networks. For example, the latest version of WizardLM, a ChatGPT-like AI system was created to follow complicated instructions.
Eight of the top 10 language models currently used in Hugging Face are variants of LLaMA 2, some are designed as traditional text generative models. Such is the impact that Lambert claims that LLaMA 2 is the biggest event of the year related to artificial intelligence. In his opinion, proprietary software has the advantage today, although he believes that new versions of LLaMA will soon be able to perform the same tasks as ChatGPT.
The researcher mentions the unanswered questions that the release of LLaMA 2 leaves for most users, due to the lack of documentation on the data used to train the model. He also assures that only major shareholders in the AI business such as Meta, OpenAI, Microsoft, or Google will have the computing resources, financial resources, and personnel to design bigger and better language models. However, Lambert is confident that, despite the success of OpenAI’s private ownership approach, the industry is entering a new era of transparency.
Recall the voluntary agreement between the White House and the seven technology companies involved in the AI business: Meta, Microsoft, OpenAI, Amazon, Alphabet, Inflectio,n and Anthropic. The agreement establishes that the companies must check aspects such as the impact on society, discriminatory potential, avoidance of abuses, or impact on national security; all before the launches. However, the agreement could be compromised by growing doubts related to the legal liability of artificial intelligence systems and pressure from various sectors of government, with many politicians fearing malicious acts using open-source models.
The researcher who leads Google’s AI development, Demis Hassabis, disagrees with the information contained in the leaked Google document; as he assures that no company will be threatened by the incursion of open-source artificial intelligence. For Hassabis, data, access to processing capacity, and talent are the key; something that big tech will continue to protect.
On the other hand, Lambert expects upgrades to improve rapidly, as the AI ecosystem will continue to expand its trajectory. For him, this new open-source AI scenario opens the door to many possibilities for start-ups and independent developers, so we can only watch as open source continues to improve.
During an interview in June, Ahmad Al-Dahle preferred not to mention the influence that the leak of the first version of LLaMA had on the creation of the new strategy for LLaMA 2. Although he announced that the company was working on the new development, which will be called LLaMA 3, he did not specify how it will be different from its predecessors.
For now, we have to wait for the company led by Mark Zuckerberg to officially launch the virtual assistants of AI Personas; as it will be a sample of the potential of the new language model and the purpose of data collection of Meta algorithms.
