Reasoning Devrim: OpenAI o1'den DeepSeek-R1'e — Test-Time Compute ve Chain-of-Thought'un Yeniden Doğuşu

Ders Haritası (10 Bölüm)#

1. Pre-reasoning era — GPT-4 sınırı
2. Chain-of-Thought — Wei 2022, prompting trick
3. o1 lansman — OpenAI Eylül 2024
4. Test-time compute scaling — Kaplan'ın yeni boyutu
5. RL for reasoning — process reward model
6. DeepSeek-R1 (Ocak 2025) — open-source breakthrough
7. Reasoning tokens — hidden (o1) vs visible (R1)
8. AIME, MATH, Codeforces benchmarks — quality leap
9. Cost economics — reasoning expensive
10. Türkçe reasoning — practical implications

1-5. Reasoning Evrim#

1.1 Pre-reasoning era (GPT-4 era)#

• Simple arithmetic: OK
• AIME problems: %12 accuracy
• Hard math olympiad: <%5

1.2 Chain-of-Thought (Wei 2022)#

Query: '23 × 47 = ?'

Kötü: '1081' (yanlış, model atlama yapıyor)

İyi (CoT prompt):
'Let me think step by step.
23 × 47 = 23 × 40 + 23 × 7
      = 920 + 161
      = 1081'

1.3 CoT'in sınırı#

1.4 o1 (OpenAI, Eylül 2024)#

• Pre-trained base + extensive RL on math/code/reasoning tasks
• Reward: final answer correctness
• Model discovers long reasoning strategies

User: 'AIME 2024 Problem 1: ...'
Model internal: '<reasoning>Let me think... try approach A... no, B...</reasoning>'
Model output: 'Answer: 42'

1.5 Test-time compute scaling#

Accuracy = f(train_compute, test_time_compute)

1.6 RL training detay#

Reward(reasoning_step) = (helpful_to_final_answer, no_logical_errors)

1.7 Self-play + search#

Thought 1 → Critic: 'wrong direction'
Backtrack → Thought 2 → Critic: 'good'
Continue...
Final answer

6-10. DeepSeek-R1 + Pratik#

6.1 DeepSeek-R1 (Ocak 2025)#

• Full paper, architecture, training recipe
• Distilled smaller variants (R1-distilled 7B, 32B)
• Open weights HuggingFace
• Quality: comparable o1

• Base: DeepSeek-V3 671B MoE (active 37B)
• RL training: GRPO (Group Relative Policy Optimization)
• Process reward + outcome reward combination
• 30M reasoning examples
• ~$5M training cost

6.2 Hidden vs Visible reasoning#

6.3 Reasoning quality leap#

• GPT-4o: %12
• Claude 3.5 Sonnet: %16
• o1-preview: %75
• o1: %83
• DeepSeek-R1: %80

• GPT-4o: %76
• o1: %94
• R1: %93

• GPT-4o: %12 percentile
• o1: %89 percentile (expert programmer)
• R1: %85

6.4 Cost economics#

• o1 input: $15/1M (vs GPT-4o$2.5)
• o1 output: $60/1M (vs GPT-4o$10)
• 4-6x more expensive

6.5 When to use reasoning models#

• Math, science, coding (high benefit)
• Multi-step planning (chain-of-thought needed)
• NOT for: simple Q&A, summarization, creative writing (overkill)

6.6 Türkçe reasoning#

• Math problem reasoning: OK (English internal, output Turkish)
• Türkçe-specific tasks (Türkçe grammar reasoning): weaker than English

6.7 Future#

• 2025-2026: reasoning standard, all major labs
• Test-time compute scaling continues
• 'Reasoning agents' — multi-LLM verification
• AI safety implications: model 'thinks' before output, more controllable

Sıradaki Ders: DeepSeek-R1 Self-Host#