South Korea is building the largest AI data center in the world

Hi pals! Here’s what you need to know about AI today:

👉 South Korea is building the largest AI data center in the world

👉 Less data with more quality is the key to build better AI

👉 Meta is scheduling its first Generative AI conference

and many more!

📧 Did someone forward you this email? Subscribe here for free to get the latest AI news everyday!

Read time: 5 minutes

What’s going on: South Korea is investing $35 billion to build the world’s largest AI data center in South Jeolla province, with a 3GW power capacity. Backed by the Ministry of Science and ICT, construction is set to start later this year and finish by 2028. The project aims to position South Korea as a global AI leader, surpassing the capacity of major tech giants’ data centers.

What does it mean: AI is unbelievably energy-hungry and those countries/companies that are able to generate insane amounts of electricity are going to win the AI race. Weeks ago, India announced its plans to build the largest data center in the world (5 times bigger than the largest Microsoft data center). Now, it’s South Korea’s turn to show off its competence in this sector. This initiative will most likely boost its economy with thousands of jobs and attracting tech investments. It could set a new benchmark for data center capacity globally as well.

More details: 

• The data center will feature advanced AI computing resources, including Nvidia H100 GPUs, and is nearly three times larger than other major facilities.

• Construction begins in 2025 and aims for completion in 2028, with support from government and private sectors.

• To facilitate the center’s construction, the government will create a special-purpose corporation (SPC) in collaboration with private enterprises, allocating an initial investment of $277 million. The SPC’s ownership structure will consist of 51% public and 49% private stakes.

What’s going on: Researchers at Shanghai Jiao Tong University have discovered that large language models can be effectively trained for reasoning tasks without requiring vast amounts of data, challenging the conventional belief that massive datasets are essential for AI development. The study suggests that strategic data selection and efficient training methods can yield impressive results even with limited resources.

What does it mean:  The key insight from the research is that the quality of data, rather than sheer quantity, plays a critical role in enabling LLMs to perform complex reasoning. By carefully curating datasets that include diverse examples and focusing on high-quality, reasoning-specific content, the models demonstrated substantial improvements in tasks like problem-solving and logical deduction. The researchers also emphasized the importance of fine-tuning techniques and iterative training processes, which allowed the models to adapt and learn more efficiently.

More details:

• In this study, they have introduced the concept of “less is more” (LIMO). Their work builds on top of previous research that showed LLMs could be aligned with human preferences with a few examples.

• A Qwen2.5-32B-Instruct model fine-tuned on 817 training examples chosen based on LIMO reached 57.1% accuracy on the highly challenging AIME benchmark and 94.8% on MATH, outperforming models that were trained on a hundred times more examples.

• The LIMO model achieved 66.7% accuracy, close to OpenAI-o1-preview’s leading score of 73.3% on the challenging GPQA benchmark.

• Interested in learning more? Check out their paper.

📰 The New York Times has approved the use of AI tools for its product and editorial staff to enhance their work.

💰 Safe Superintelligence, an AI startup founded by former OpenAI chief scientist Ilya Sutskever, is close to raising more than $1 billion at a $30 billion valuation.

🦙 Meta has announced its first generative AI developer conference, LlamaCon, scheduled for April 29, 2025, where it will share updates on its open-source AI developments to help developers create innovative apps and products. Signup here to get notified about LlamaCon updates.

⚖ OpenAI is considering granting special voting rights to its nonprofit board to prevent future hostile takeover attempts, including a recent $97.4 billion offer from Elon Musk.

🖥 Fiverr, a famous freelancer marketplace, has introduced new AI tools under the "Fiverr Go" brand, allowing freelancers to train AI models on their past work, such as voice-overs and graphic designs, and charge clients for access, aiming to help them scale their services.

🚀 Former OpenAI CTO Mira Murati has launched a new AI startup called Thinking Machines Lab, aiming to create more customizable and widely understood AI systems.

AI + Writing work emails

Perplexity - Technical Data Annotation Analyst

Nokia - Decentralized Medical AI Research Intern

Faire - Senior Data Scientist

Xometry - Data Science Intern

Thank you for staying with us like always! If you are not subscribed, subscribe here for free to get more of these emails in your inbox! Cheers!