The YZ50 program, aiming to move Turkey’s artificial intelligence ecosystem from a consumer of ready-made APIs to a producer position developing fundamental models, has been announced. The initiative, which will select 50 talented young people for a 12-week intensive research camp, offers participants an in-depth curriculum ranging from first-line neuron coding to building their own small GPT models from scratch.

Moving Beyond Using Ready-Made APIs: 50 Young People Developing Models

With the recent proliferation of generative artificial intelligence tools, a large portion of the sector is focusing on integrating ready-made language models via interfaces or APIs. The YZ50 program, however, is positioned in direct opposition to this approach. The initiative’s core ideology rests on the thesis that an AI researcher must build the model from scratch, with their own hands, in order to truly understand it.

Centered around the famous

“Zero to Hero” methodology of former OpenAI co-founder and Tesla AI director Andrej Karpathy, the program targets candidates with a background in mathematics, computer science, and engineering. For 12 weeks, participants not only watch theoretical lectures or videos; they also submit working code blocks, experiment logs, and technical presentations every week.

Program Structure and Working Models

YZ50 offers two different tracking paths depending on the candidates’ time management:

• Full Track: The core team that can dedicate 4 to 6 hours of intensive work per day, submitting code, experiment logs, and technical presentations every week. Participants in this group are directly involved in the 60-minute defect analysis and demo meetings held every Friday at 5 PM.
• Light Track: A group that can dedicate an average of 10 hours per week, following the technical curriculum at a more flexible pace, progressing through monthly checkpoints.

12-Week Roadmap: From Neuron to Your Own GPT System

The The training process consists of three main phases. Before participants use well-known deep learning libraries as ready-made functions, they personally code the mathematical operations that these libraries perform in the background.

Weeks

Scope and Outputs

Weeks 1-4

Neuro simulation, writing an autograd engine from scratch with micrograd architecture, PyTorch fundamentals and character-level language modeling (makemore).

Weeks 5-8

Embedding layers, hyperparameter optimization, Kaiming weight initialization, manual matrix-level backpropagation and WaveNet Architecture.

Weeks 9-12

Self-attention system, building Transformer blocks from scratch, pre-training/fine-tuning processes, and a small working GPT capstone project.

Global Academic Staff Provides Mentorship

The program’s researcher and mentor team includes Turkish scientists conducting doctoral/research studies on deep learning and fundamental models at institutions such as Stanford, UC Berkeley, University of Pennsylvania (UPenn), and Bilkent AI. Researchers such as Tunga Bayrak, Batu El, Mert Cemri, and Yavuz Öztürk directly guide participants through the error debugging processes and technical reporting.

A Critical Step for Turkey’s Deep Technology Future

In an era where large language models and artificial intelligence infrastructures are transforming into a global dominance race, the development of intellectual capacity in Turkey that understands these systems at the architectural level is of strategic value. YZ50 aims to transform its participants not only into implementers but also into researchers who can contribute to open-source models on a global scale or train local fundamental models. The open-source small GPT applications and technical reports that will emerge at the end of the process promise to directly contribute to the domestic artificial intelligence literature.