The Intersection of Innovation, Privacy, and Collaboration in South Korea's Tech Landscape

I seized the opportunity to reconnect with some familiar faces from last year's trip - Pete Tae-hoon Kim and Hyun-Kyu (Logan) Lee from Deeping Source Inc. . Pete, with his insightful leadership as CEO, and Logan, whose expertise continually drives the company forward, have become more than just acquaintances; they are friends who embody the spirit of innovation and camaraderie I've come to associate with South Korea.

Our reunion started with a delightful lunch, where conversations flowed as effortlessly as the Han River, weaving through topics of technology, culture, and personal growth. The food, a gastronomic representation, was as delightful as the company. I was reminded of this country's unique warmth and hospitality as we shared stories and laughed.

Post-lunch, the anticipation built up as we headed towards the Deeping Source office. I was eager to witness first-hand their developmental strides since my last visit. A thought had been simmering in my mind, particularly in light of recent events like The Times' lawsuit against OpenAI and Microsoft concerning AI's use of copyrighted work. With the rapid advancement and versatility of multimodal AI.

This led me to ponder:

Can we harness the rich data from video footage and real-time streams to enhance our AI models while upholding the utmost respect for personal privacy?

It's a question that strikes at the heart of achieving 'Superhuman AI' as I touched upon in the write-up up 7 Key Insights and Predictions for AI in 2024

Current large language models (LLMs) require an enormous volume of text data for training, equivalent to what would take a human 20,000 years to read. Despite this extensive training, these models still struggle with basic logical concepts, such as understanding that if A equals B, then B equals A. This highlights a significant gap between AI and biological intelligence.

In contrast, humans and animals demonstrate remarkable learning efficiency. For example, humans, with their complex brains, and animals like corvids, parrots, dogs, and octopuses, with far fewer neurons, exhibit rapid learning with minimal data. This ability suggests that the architecture of biological brains is vastly more efficient than current AI systems.

The future of AI might lie in developing new architectures that mirror the learning efficiency seen in biological entities. Relying solely on increasing text data, synthetic or otherwise, is likely just a temporary solution, given the limitations of current methodologies. A more promising approach might involve integrating sensory data, such as video, into AI training. Video data offers higher bandwidth and more structured information compared to text, presenting an opportunity to build more sophisticated and efficient AI systems.

For perspective, consider the amount of visual data a two-year-old child processes. In about 32 million seconds (equivalent to two years), a child’s optical nerves, with around 2 million fibers transmitting approximately ten bytes per second, would process around 6E14 bytes of data. This volume is significantly higher than the 2E13 bytes used in typical LLM training. Importantly, visual data, being more redundant, offers richer insights into the structure of the world than text data. It encompasses a broader range of learning stimuli, from spatial relationships to dynamic interactions, which are crucial for developing more advanced, efficient AI systems.

This goes hand in hand with Sam Altman's talk at Y Combinator W24 kickoff, where he provided insights into the anticipated features of GPT-5, particularly emphasizing its multi-modal capabilities. Here are the key points:

1. Multi-modality with a Focus on Video: Altman highlighted the significant role of video in the upcoming GPT-5. He believes that integrating video processing and output will be a major advancement. This suggests that GPT-5 could handle inputs and outputs in video format, a leap forward in AI interaction.
2. Enhanced Reasoning Abilities: Altman noted that while GPT-4 has limited reasoning capabilities, comparable to a child's, GPT-5 aims to improve in this area markedly. The expectation is for GPT-5 to approach complex situations more thoughtfully, possibly unlocking advanced cognitive abilities akin to human 'System 2' thinking - deliberate and analytical.
3. Reliability and Consistency: A critical focus for GPT-5 is reliability. Altman pointed out that while GPT-4 might produce a good response in one out of 10,000 attempts, it cannot identify which one is the best. GPT-5 aims to overcome this by learning and remembering the most effective ways to perform tasks, thereby offering consistent and reliable outputs.
4. Customization and Personalization: Altman also stressed the importance of customization and personalization in AI. He suggested that GPT-5 would be designed to adapt to various user styles and preferences, integrating personal data sources like emails or calendars for more tailored interactions.

Additionally, in an interview with Bill Gates, Altman expanded on these points, particularly the multi-modal aspect. He envisaged a future where AI, like GPT-5 leveraging extensive datasets of labeled videos for training. These advancements hint at a future where AI will excel not only in text and speech but also in understanding and generating video content, moving closer to a more comprehensive form of general intelligence. This progression underlines AI technology's rapid and ongoing evolution as we move into 2024.

Now let's explore the later part of my initial question:

Can we harness the rich data from video footage and real-time streams to enhance our AI models while upholding the utmost respect for personal privacy?

This is what I love about Deeping Source's technology: it primarily revolves around advanced AI video analytics combined with a robust data anonymization system. Their AI solution anonymizes video data by removing Personally Identifiable Information (PII) while retaining essential features for analysis, such as age, gender, facial expressions, and posture. This approach allows for detailed analytics without compromising individual privacy.

SEAL is an advanced AI-powered solution for anonymizing images and videos while preserving essential AI features. It efficiently removes personal identifiers like faces and license plates, ensuring privacy. Unique for its ability to maintain AI data utility, SEAL supports real-time processing and works well with real-world data. It's designed for easy integration and is suitable for various AI vision applications. SEAL emphasizes balancing rigorous privacy standards with enhanced AI performance, offering deployment options such as SaaS, API, embedded, and on-premise solutions.

Deeping Source's technology finds applications across sectors like retail, logistics, banking, smart cities, and cultural spaces. In retail, it helps in understanding customer behavior and preferences. In logistics and industry, it enhances safety and efficiency. For banking, it aids in detecting suspicious activities at ATMs. In smart cities, the technology assists in large-scale space management and simulation.

The company's approach to integrating AI and anonymization technology is particularly significant in the context of privacy concerns in the digital age. By ensuring that data collection and analysis are privacy-safe, Deeping Source positions itself as a solution that balances the need for in-depth analytics with the imperative of protecting individual privacy rights.