# Kara Büyünün Ardında | Burak Sina Akbudak | TEDxIzmir Fen Lisesi Youth

Source: https://www.youtube.com/watch?v=Z7XxjboR7tY
Recap page: https://rapidrecap.app/video/Z7XxjboR7tY
Generated: 2025-12-22T16:28:09.285+00:00

---
## Quick Overview

The presentation by Burak Sina Akbudak at TEDxIzmir Fen Lisesi Youth focuses on demystifying Artificial Intelligence, particularly Large Language Models (LLMs), by outlining the key steps in their creation—Data Sets, Model Architecture, Training, Validation, Extra Precautions, and Inference—while emphasizing the accelerating pace of development and the need for humans to adapt to AI as a powerful new tool, rather than something entirely foreign.

**Key Points:**
- The speaker, Burak Sina Akbudak (aka cos beta), is currently studying Computer Engineering and has experience in NLP projects and competitions like ZOOAA Turkey.
- AI development is structured in clear phases: Pre-training (using massive data sets like Common Crawl to teach language basics), Fine-Tuning (specializing the model for specific domains via labeling), and RLHF (Reinforcement Learning from Human Feedback) to align outputs with human preferences.
- The scale of data processing is immense; Common Crawl adds 250 TB of data monthly (2 billion web pages) since 2015, accumulating petabytes of data.
- Model complexity is demonstrated by the rapid growth in parameters, moving from the 13-year-old Human model (<100 Million tokens) to DeepSeek-V3 (2024) processing 14.8 Trillion tokens.
- Practical AI applications span Natural Language Processing (code generation, Q&A), Computer Vision (object tracking, pose estimation, image generation via Diffusion/GANs), and Audio/Music processing (synthesis, source separation, classification).
- The core message for the future is that 'Tools change, humanity adapts' (Araçlar değişir, insanlık adapte olur), comparing AI access to the progression from libraries to the internet to AI itself.
- The training process, especially for LLMs, requires massive computational resources, exemplified by the Grok-3 model needing 200,000 GPUs, costing $5 billion.

![Screenshot at 00:41: The speaker points to the high-level overview of the AI development process, detailing the sequential steps: Data Sets -\> Model Architecture -\> Training -\> Validation -\> Extra Precautions -\> Inference, establishing the structure for the rest of the technical discussion.](https://ss.rapidrecap.app/screens/Z7XxjboR7tY/00-00-41.jpg)

**Context:** Burak Sina Akbudak delivered this talk at TEDxIzmir Fen Lisesi Youth, focusing on explaining the underlying mechanics and practical applications of modern Artificial Intelligence, especially Large Language Models (LLMs). The presentation systematically walks the audience through the lifecycle of building an AI model, from initial data collection and architecture design to advanced fine-tuning and validation techniques, using analogies like architecture (Mimari) versus the final product (Model) to clarify complex concepts.

## Detailed Analysis

The presentation systematically breaks down the world of modern AI, focusing heavily on Large Language Models (LLMs). Akbudak begins by introducing himself, highlighting his background in computer science and NLP projects. He then moves to foundational definitions, explaining the hierarchy of AI, Machine Learning (ML), Deep Learning (DL), and LLMs as concentric circles, and defining key terms like AGI, Parameters, and Tokens. The talk transitions into the AI development pipeline, which involves Pre-training on massive datasets like Common Crawl (250 TB monthly since 2015), Fine-Tuning for specialization, and RLHF (Reinforcement Learning from Human Feedback) for alignment, illustrated by a complex flowchart. He emphasizes the exponential growth in model size, citing figures from 13 y.o. Human (<100 Million tokens) to DeepSeek-V3 (14.8 Trillion tokens). Practical applications are categorized into NLP (Q&A, coding), Computer Vision (image generation, object tracking), and Audio/Music (synthesis, classification). The talk concludes by framing AI as the next evolutionary tool for humanity, similar to the transition from libraries to the internet, urging the audience to adapt to this new era where 'Tools change, humanity adapts.' Specific examples of the massive infrastructure required, such as Grok 3 needing 200,000 GPUs at a $5 billion cost, underscore the scale of current AI research.

### Introduction and Definitions

- Speaker introduces himself (Burak Sina Akbudak, aka cos beta)
- Overview of AI hierarchy (AI contains ML, which contains DL, which contains LLM)
- Definitions of AGI, Parameters, and Tokens provided.

### Data Sets (Pre-training)

- Models learn language intricacies from vast datasets like Common Crawl (250 TB/month) and sources like Wikipedia, Reddit, and Pinterest
- Emphasis on the sheer volume ('Çok büyük veri') required for initial training.

### Model Architecture vs. Model

- Architecture (Mimari) is compared to a blueprint (listing Transformers, Diffusion, GANs, CNNs, LSTMs, SVMs), while the Model is the resulting product (listing GPT-4o, DALL-E, Suno, BARK, BERT, ResNet).

### Training Pipeline (Eğitim)

- AI training follows a sequence: Pre-training -> Instruction Tuning -> Reward Modeling -> RL Update (Reinforcement Learning) -> Validation -> Extra Precautions -> Inference.

### Scale of Growth (Kısa Bir Tarih Dersi)

- Model size has grown exponentially, from <100 Million tokens (2018) to 14.8 Trillion tokens (DeepSeek-V3, 2024)
- Training Grok 3 required 200,000 GPUs at a $5 billion cost.

### Practical AI Applications (Pratikte AI)

- Divided into NLP (code writing, Q&A), Computer Vision (image generation, object tracking, pose estimation), and Audio/Music (synthesis, source separation, classification).

### Conclusion (AI'ı Nasıl Düşünmelisiniz)

- AI is a tool that humanity adapts to, following historical shifts in information access (Library -> Internet -> AI) and task completion (Family -> Friend -> Self).

![Screenshot at 00:00: Establishing shot of the TEDx stage with the 'TEDx Izmir Fen Lisesi Youth' banner and the speaker approaching the podium.](https://ss.rapidrecap.app/screens/Z7XxjboR7tY/00-00-00.jpg)
![Screenshot at 00:41: Slide displaying the core structure of AI training/development steps: Data Sets, Model Architecture, Training, Validation, Extra Precautions, and Inference.](https://ss.rapidrecap.app/screens/Z7XxjboR7tY/00-00-41.jpg)
![Screenshot at 02:14: Slide detailing key definitions in AI, showing the nested relationship between AI, Machine Learning, Deep Learning, and LLMs, alongside definitions for AGI, Parameter, and Token.](https://ss.rapidrecap.app/screens/Z7XxjboR7tY/00-02-14.jpg)
![Screenshot at 04:44: Slide illustrating the exponential growth of AI models over time based on the number of tokens seen during training, spanning from 2018 to 2024.](https://ss.rapidrecap.app/screens/Z7XxjboR7tY/00-04-44.jpg)
![Screenshot at 13:13: Slide contrasting 'Extra Precautions' \(like preventing harmful outputs, shown with GPT-4 responses\) with 'Inference' \(model use, shown with interface examples\).](https://ss.rapidrecap.app/screens/Z7XxjboR7tY/00-13-13.jpg)