Meta making 'tremendous progress' towards an AI future

Meta announced at a Seoul briefing that its Llama AI model has become more efficient and versatile while showcasing collaborations with local startups for enhanced applications.

Meta Vice President Manohar Paluri speaks during a media briefing at Meta’s Korean headquarters in Gangnam District, southern Seoul, on Thursday. [CHO YONG-JUN]

Meta told Korean press that its open source language model Llama, which has given birth to 65,000 derivative AI models, is more efficient than ever at a briefing in Seoul on Thursday. It caveated, however, that it still has a long way to go reaching the AGI dream.

The company will continue to press on, Meta Vice President Manohar Paluri said at the event at Meta's Korean headquarters in Gangnam District, southern Seoul, but will “do it in a way that is responsible and bring along everybody in the world with you so the transformation is affordable and helpful for everyone.”

"From the point of view of over the last decade with the acceleration of AI, we are actually making really good progress to unlocking [AGI] models," he said. "It's very heard to measure [in quantity] but we are definitely making a tremendous amount of progress."

AGI is a hypothetical type of AI with humanlike intelligence and the capability to teach itself. Compared to conventional AI systems, wherein models are customized for specific use cases — a language model, for example, or a text-to-image generator — AGI would ideally have its cognition to combine such systems and process on its own.

Meta released Llama 3.1, the latest version of the company’s open-source AI model, in July, claiming that it outperforms OpenAI while being more efficient than competition models. Llama 3.2, announced at Meta Connect 2024 last month, is its first open source model capable of processing both images and text. The company used a customized version of Llama to create Meta AI, which is implemented across the company’s platforms including Facebook, Instagram and WhatsApp.

“We are making really good progress toward unlocking these models working out of the box for many use cases,” the vice president said, adding that it would be very difficult to measure that progress with a percentage.

Meta claimed that it made Llama open source as it first wanted more people to use and leverage the technology first.

“We have had a history of first breeding things to making them work at scale and then thinking about monetization,” the vice president said.

“As we start making progress with the AI products there will be ways to monetize for it.”

AI startup Aim Intelligence and the Korea Institute of Science and Technology of Information (KISTI) joined the media briefing and demonstrated their use cases.

Aim Intelligence, an AI cybersecurity startup, fine-tuned Llama and its safeguard tool, Llama Guard 3 8b, with the firm’s Korean-specific data to create Llama Suho, a content filtering model that is more efficient for use in Korea.

KISTI created KISTI Open Natural Intelligence (KONI), a science and technology-focused large language model based on Llama and trained on its large database.

BY CHO YONG-JUN [cho.yongjun1@joongang.co.kr]