In an attempt to outperform OpenAI, Meta Platforms has made early versions of its most recent large language model, Llama 3, available.
In an attempt to overtake OpenAI, the leader in the generative AI market, Meta Platforms has made early versions of two of its most recent large language models available: Llama 3 and an image generator that updates images in real time as users write prompts.
The models will be incorporated into the company’s virtual assistant, Meta AI, which it is marketing as the most advanced of its free-to-use competitors. Performance comparisons on topics like coding, reasoning, and creative writing are used to highlight Meta AI’s superiority over products from competitors like Google and French start-up Mistral AI.
In order to better compete with Microsoft-backed OpenAI’s breakthrough smash, ChatGPT, the revised Meta AI assistant will receive more prominent billing within Meta’s Facebook, Instagram, WhatsApp, and Messenger apps in addition to a new standalone website.
On that website, a welcome landing page invites users to try hiring an assistant to help with homework, make a packing plan for their vacation, play music trivia from the 1990s, and draw pictures of the skyline of New York City.
In an effort to challenge OpenAI’s dominant position in the market for generative AI, Meta has been rushing to provide these products to its billions of users. This has involved a costly revamp of computing infrastructure and the merging of formerly separate research and product teams.
As part of its attempt to catch up, the social media behemoth has been freely distributing its Llama models to developers creating AI apps. This potent free option has the potential to thwart competitors’ ambitions to make money from their exclusive technology. Critics of the approach are concerned about safety because they are apprehensive about what dishonest people might utilize the concept to create.
Llama 3 now has new computer coding capabilities according to Meta, which also fed it images and text during training. However, Chris Cox, chief product officer at Meta, stated in an interview that the model will only output text at this moment.
Differentiation
Subsequent versions will have more sophisticated reasoning, such as the capacity to create lengthier multi-step plans, he said. According to blog entries from Meta, versions that are scheduled for release in the upcoming months will also possess “multimodality,” or the ability to produce both text and graphics.
“The ultimate aim is to alleviate burdens from you and simplify your life, be it arranging travel, composing written assignments, or engaging with businesses,” Cox stated.
According to Cox, the Ray-Ban Meta smartglasses, a product of Essilor Luxoticca, will benefit from an update this year that includes picture training for Llama 3. This update will allow Meta AI to recognize items viewed by the wearer and respond to inquiries about them.
Late Thursday saw a 1.8% increase in Meta shares.
In addition, Meta announced a relationship with Google to enhance its current agreement with Microsoft’s Bing search engine by including its real-time search results into the assistant’s responses.
With the update, the Meta AI assistant is now available in over a dozen non-US markets, including Australia, Canada, Singapore, Nigeria, and Pakistan. According to Cox, Meta is “continually figuring out how to do this in Europe,” where privacy laws are more strict and the upcoming AI Act is set to impose regulations like disclosing model training data.
The insatiable need for data that generative AI models exhibit has become a major source of friction in the advancement of the technology.
In a video that accompanied the announcement, Meta CEO Mark Zuckerberg referred to Meta AI as “the most intelligent AI assistant that you can freely use” and gave a hint to the competition with OpenAI.
According to Zuckerberg, on performance measures that are frequently used to evaluate model quality, the two smaller Llama 3 versions that are currently being released, with eight billion and seventy billion parameters, performed favorably when compared to other free models. With 400 billion parameters, the largest version of Llama 3 is still being trained, he said.
“Undoubtedly impressive,” those results also pointed to a widening performance disparity between proprietary and free models, according to Nathan Benaich, the founder of AI-focused startup capital firm Air Street Capital.
The previous Llama 2 model, according to its developers, confused requests for directions on how to “kill” a computer program with those for murderous methods since it was unable to comprehend basic context. rival Similar issues have been encountered by Google, which recently suspended the usage of its Gemini AI image production tool following complaints that it was producing erroneous portraits of historical personalities.
According to Meta, Llama 3’s issues were reduced by employing “high-quality data” to train the model to recognize subtleties. Although it claimed to have fed seven times as much data into Llama 3 as it had Llama 2, it did not go into detail about the datasets used.
Leave a Reply