🤖 Hello, this is CLEVI.
From 20 to 22 January, I attended the 「2025 JBNU Research Fair Week」 event hosted by Jeonbuk National University.
Centred on Jeonbuk National University’s outstanding research achievements, this event was a meaningful occasion that presented the direction of future technologies linked to industrial and national strategies and provided a forum for exchange involving industry, academia, research institutes and government.
In particular, CLEVI CEO Lee Hwan-ho was invited as a VIP speaker at the opening ceremony held on the 20th! 👏
Let’s relive the atmosphere of that day!
This is CEO Lee Hwan-ho preparing his talk before the event began. You can see him thinking about how to make the talk more informative and accessible.
The event began with the opening ceremony.
The event opened in spectacular fashion with a live art performance completed spontaneously by a Jeonbuk National University student.
Following the first keynote speech, CLEVI’s CEO Lee Hwan-ho gave the second keynote speech on the topic, “Why Does AI Need a Physical Medium?” Shall we take a brief look?
How exactly does AI understand the world,
create sentences and even take action?
We are going to look inside the ‘brain’ of large language models (LLMs), such as ChatGPT and Claude, which we use every day, to understand exactly how they understand and create sentences.
- Giving Names (Token ID): AI first manages all words by assigning each one a unique number, such as ‘5’ or ‘8’.
- Pinning Addresses on a Map (Vector): However, because a simple number cannot convey meaning, each word is assigned a ‘detailed address (coordinate)’ made up of hundreds of numbers.
- Finding Neighbours: Words with similar meanings (e.g. ‘the’ and ‘a’) live close together on this map, like neighbours.
So how does AI predict the next word?
Now that we understand basic data transformation, let’s look at the entire process of how these numbers become an actual sentence. Let’s assume that we have an input here: "Paris is the city..."
Tokenizer (breaking into pieces): As we saw earlier, the sentence we enter is converted into numbers (Token IDs) through the tokenizer. "Paris" is number 1652, "is" is number 5... and so on.
LLM (a giant calculator): These clusters of numbers enter a giant calculator called an LLM (the Bloom model shown in the image). Here, it understands the context through the 'vector' operations we learned about earlier. 'Ah, we’re talking about the city of Paris right now!' and so on.
Logits & Softmax (calculating probabilities): Once it has finished calculating, the AI scores tens of thousands of word candidates that could come next. It then passes these through a filter called 'softmax' to convert them into **'probabilities'** that are easier for us to understand.
Final selection: Look at the graph on the right.
- of → 33.7%
- that → 15.3%
- where → 9.4%
Ultimately, the AI selects **'of'**, which has the highest probability, and completes the sentence "Paris is the city of...". It repeats this process tens of thousands of times to create the long answers we see.
In conclusion, for AI, remembering a word is not about memorising the letters; it is like remembering the 'address' of where that word lives on a vast map of meaning.
LLMs can be extended with vision, robotics and behavioural learning.
AI that understands meaning, sees and understands, and acts.
Giving AI arms and legs means extending intelligence from language to perception, and further to action.
AI is no longer in the realm of ideas or imagination; it has entered the stage of implementation and reality.
CLEVI’s next step
CLEVI plans to continue expanding various opportunities, including on- and off-campus workshops, hands-on projects and experience booths, so that more people can directly experience how AI technology works in the real world.
Just as this lecture at Jeonbuk National University became a valuable bridge connecting the students and us, we will continue to serve as a bridge between engineers and users, and between academia and industry.