[AI One-Bite News] Even Zuckerberg Joins In... Nvidia Smiles in the AGI War

Zuckerberg "Will Release AGI as Open Source"
Secured 350,000 H100 Units... Large-Scale Infrastructure Investment

Editor's NoteSince the emergence of ChatGPT, the world has been buzzing with artificial intelligence (AI). Technology is changing rapidly, and news is pouring in every day. Even top industry experts are struggling to keep up. It's overwhelming to consume it all, but you can't afford to miss out either. We deliver the core of weekly AI issues, carefully selected and easy to digest.

Meta Joins the AGI Race... Betting 30 Trillion Won on Infrastructure

Mark Zuckerberg, CEO of Facebook

Following OpenAI, Meta is challenging Artificial General Intelligence (AGI). AGI is known as the "dream technology" that possesses intelligence comparable to or on par with humans. Mark Zuckerberg, CEO of Meta, stated on his Facebook on the 19th, "The long-term vision is to provide general intelligence as open source so that everyone can benefit." This means taking a different path from OpenAI, which does not disclose its technology. Zuckerberg also announced plans to secure more than 350,000 of NVIDIA's high-performance AI chips, the 'H100,' by the end of this year for the future roadmap. Including other graphics processing units (GPUs), Meta will have infrastructure equivalent to 600,000 H100 chips. Considering that each H100 costs about 50 million won, this amounts to an investment of 30 trillion won. In an interview with IT media The Verge, Zuckerberg said, "We have concluded that we need to create products for general intelligence," adding, "Making it open can solve various problems that may arise from unequal access to opportunities and value."

Geometry Made Easy... AI Aiming for Gold at the Math Olympiad

Google DeepMind and New York University joint researchers have developed an AI that solves math problems at the level of the International Mathematical Olympiad. The AI, named 'AlphaGeometry,' can solve geometry problems with abilities surpassing humans. Geometry has long been a 'fourth-dimensional wall' for AI, as it requires advanced reasoning skills and it was difficult to create training data. There are few human-solved geometry problems, and the various symbols used in geometry are hard to teach. To address this, the researchers trained the AI on over 100 million geometric proofs generated algorithmically rather than from human solutions. As a result, AlphaGeometry solved 25 out of 30 International Mathematical Olympiad problems, a level comparable to past gold medalists.

'Copyright-Free AI' Certification Introduced

"We certify AI that does not infringe copyright." A certification has been introduced for AI that resolves copyright issues. It serves to reassure users so they do not worry about legal risks. To obtain certification, AI providers must submit the sources of their training data and related details, and undergo an evaluation of data usage consent. In addition to fees paid to the certifying company, they must pay up to $6,000 annually (about 7.8 million won) proportional to their revenue. So far, nine generative AI startups have received certification. The certifying company was founded by a former Stability AI executive. Ironically, this is the same company known for developing the image-generating AI 'Stable Diffusion,' which has been involved in numerous copyright lawsuits.

Beyond Faces and Voices... AI Mimicking Human Handwriting

Following faces and voices, AI that mimics human handwriting has emerged. A research team at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi developed an AI capable of imitating handwriting with just a few paragraphs of text. In tests involving 100 people, the team found that the original handwriting and the AI-generated imitation were almost indistinguishable. This technology can help people with physical disabilities write without holding a pen, but it also poses risks of forgery and misuse. The research team is currently seeking partners to commercialize the technology.