Subtotal 0,00 د.إ Subtotal: 0,00 د.إ
RISE MODEL
The Rise of Reasoning Models
"Similar to how a human may think for a long time before responding to a difficult question, o1 uses a chain of thought when attempting to solve a problem."
OpenAI, Learning to Reason with LLMs
For the first time in AI history, models aren’t just good at pattern recognition — they’re being trained to use their mind well.
AHA
The "Aha" Moment That Changed Everything
September 2024. OpenAI quietly dropped a model that answered a question most engineers had been asking for two years: what if instead of just training bigger models, we let them think longer?
o1 didn’t just answer questions. It reasoned through them. It checked its own work. It tried a different approach when the first one failed. And it did this with a kind of transparency that felt genuinely different from the autocomplete-on-steroids that LLMs had been.
We’re now in the era of reasoning models — and it’s reshaping how we build AI systems.
REASONING
What Is a Reasoning Model?
September 2024. OpenAI quietly dropped a model that answered a question most engineers had been asking for two years: what if instead of just training bigger models, we let them think longer?
o1 didn’t just answer questions. It reasoned through them. It checked its own work. It tried a different approach when the first one failed. And it did this with a kind of transparency that felt genuinely different from the autocomplete-on-steroids that LLMs had been.
We’re now in the era of reasoning models — and it’s reshaping how we build AI systems.
ACTUALLY
Under the HoodHow o1 Actually Works
o1 was trained with a new optimization algorithm combining:
- Fine-tuning on reasoning examples (supervised learning on chain-of-thought traces)
- Reinforcement learning that rewards correct final answers — not just plausible-sounding next tokens
"Through reinforcement learning, o1 learns to hone its chain of thought and refine the strategies it uses. It learns to recognize and correct its mistakes. It learns to break down tricky steps into simpler ones. It learns to try a different approach when the current one isn't working."
OpenAI, Learning to Reason with LLMs
Reverse-engineering research identified 6 steps in o1’s reasoning process — PromptLayer, 2027
Step 1: Problem Reformulation
→ Restate the problem, identify key constraints
Step 2: Decomposition
→ Break complex problem into manageable sub-problems
Step 3: Strategy Selection
→ Identify multiple solution paths
Step 4: Execution
→ Work through each step with self-monitoring
Step 5: Error Detection
→ Identify mistakes in reasoning chains
Step 6: Refinement
→ Revise and produce final answerThis mirrors how a good engineer debugs: define the problem, hypothesize, test, backtrack, fix.
PAY ATTENTION
Build Your Own VoiceThe Numbers That Made Everyone Pay Attention
This mirrors how a good engineer debugs: define the problem, hypothesize, test, backtrack, fix.
- GPT-4o: solved 12% of problems (1.8/15)
- o1 (single sample): solved 74% (11.1/15)
- o1 (best of 64 samples): solved 83% (12.5/15)
- o1 (reranking 1000 samples): solved 93% (13.9/15)
OpenAI, Learning to Reason with LLMs”
o1 also ranked 49th percentile in the 2024 International Olympiad in Informatics — under actual competition rules. That’s not a benchmark trick. That’s competitive performance among the world’s best young programmers.
BITTER
The "Bitter Lesson" This Proves
Computer scientist Richard Sutton’s “Bitter Lesson” (2019) argued that methods leveraging more compute reliably outperform clever human-designed solutions over time.
Reasoning models are the latest proof:
“The development of reasoning models illustrates Richard S. Sutton’s ‘bitter lesson’ that scaling compute typically outperforms methods based on human-designed insights.”
Wikipedia, Reasoning Model
Researchers who tried to replicate o1 using sophisticated techniques like Monte Carlo tree search found that simply having the model write out long chains of thought — and training on that — worked surprisingly well.
DEEPSEEK-R1
The Open-Source AnswerDeepSeek-R1
China’s DeepSeek released R1 in early 2027 — a reasoning model that matched o1’s performance on many benchmarks, released openly.
“Some of the most important AI advances in 2024 were test-time reasoning LLMs, or large reasoning models (LRM) — LLMs trained to write down and reuse their chains of thought for future reference.”
Synthesis AI, 2027
This opened reasoning-model capabilities to teams who couldn’t afford OpenAI’s o1-pro pricing ($150/million input tokens, $600/million output tokens — Wikipedia, OpenAI o1).
WHEN
The Open-Source AnswerWhen Should You Use a Reasoning Model?
Not every task benefits from extended thinking. Reasoning models are slower and more expensive per token than standard models.
Use Reasoning Models For
- Multi-step math or logic problems
- Complex code debugging and generation
- Legal/contract analysis
- Scientific research synthesis
- Strategic planning with multiple constraints
- Any problem where "check your work" matters
Skip Reasoning Models For
- Simple Q&A
- Summarization
- High-volume, latency-sensitive applications
- Creative writing (reasoning models can be "drier")
- Real-time chatbots
o1-preview was found to be not preferred on some natural language tasks, suggesting it’s not universally better — OpenAI.
THROUGH API
Practical CodeUsing Reasoning Models via API
# Using o3 (OpenAI's latest reasoning model)
from openai import OpenAI
client = OpenAI()
# o3 supports "reasoning_effort" parameter
response = client.chat.completions.create(
model="o3",
messages=[
{
"role": "user",
"content": """
I have a distributed system with 3 microservices. Service A calls B,
B calls C. Under load, 5% of requests to B timeout. C is healthy.
My p99 latency has doubled. Diagnose and provide a resolution plan.
"""
}
],
reasoning_effort="high" # low | medium | high — tradeoff: cost vs depth
)
print(response.choices[0].message.content)
# The model will internally "think" before answering# Claude's Extended Thinking (Anthropic's reasoning approach)
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-opus-4-6",
max_tokens=16000,
thinking={
"type": "enabled",
"budget_tokens": 10000 # tokens allocated for thinking
},
messages=[{
"role": "user",
"content": "Analyze the time complexity of this algorithm and suggest optimizations..."
}]
)
# Response contains thinking blocks + final answer
for block in response.content:
if block.type == "thinking":
print("Model's reasoning:", block.thinking)
elif block.type == "text":
print("Final answer:", block.text)FOR YOU
For Business LeadersWhat Reasoning Models Mean for You
Think of it this way: you now have access to an analyst who doesn’t just look up information but actively works through problems — checking their logic, identifying gaps, and arriving at more defensible conclusions.
Applications where this matters most
- Financial modeling — complex scenario analysis with multiple variables
- Contract risk assessment — identifying non-standard clauses across long documents
- Engineering architecture review — evaluating tradeoffs in system design
- Strategic planning — working through competitive scenarios with constraints
The tradeoff: reasoning models cost more per query and take longer. They’re tools for high-value problems, not bulk processing.
SCALING
What Comes NextThe Scaling Frontier
“OpenAI’s o1 family made this approach available at scale in September 2024 and popularized the label ‘reasoning model’ for LLMs that deliberately think before they answer.”
Wikipedia, Reasoning Model
The next frontier is adaptive reasoning — models that dynamically decide how much to think based on problem difficulty. A simple question gets a fast answer. A complex one triggers deep reasoning automatically, without you specifying.
Research directions include
- Process Reward Models (PRMs) — training that rewards correct intermediate steps, not just final answers
- Monte Carlo Tree Search at inference time — exploring multiple reasoning branches and selecting the best
- Multi-agent reasoning — multiple model instances cross-checking each other's work
Reasoning models represent the most significant capability leap since GPT-4. The ability to allocate more inference compute to hard problems — and have the model actively check its work — has unlocked performance on tasks that were previously unreliable with standard LLMs.
For engineers: understand when to reach for a reasoning model vs. a standard one. For business leaders: recognize that your AI’s ability to work through complex decisions — not just retrieve information — is now a real product differentiator.
Explore project snapshots or discuss custom web solutions.
It is not enough to have a good mind; the main thing is to use it well.
René Descartes,
Discourse on the Method
Thank You for Spending Your Valuable Time
I truly appreciate you taking the time to read blog. Your valuable time means a lot to me, and I hope you found the content insightful and engaging!
FAQ's
Frequently Asked Questions
Chain-of-thought (CoT) is a *prompting technique* — you ask the model to "think step by step." Reasoning models are trained to do this automatically using reinforcement learning. The model has internalized good reasoning as a behavior, not just a response to instructions — Simon Willison, 2024.
No. o1-preview was not preferred over GPT-4o on natural language and creative tasks. Use reasoning models for problems where logical correctness matters more than fluency or creativity.
o1-pro API: $150/million input tokens, $600/million output — significantly more expensive than standard models. Budget accordingly — factor in that fewer reasoning model calls may replace many standard model calls for complex tasks.
Yes — DeepSeek R1 and its distilled variants are available via Ollama and Hugging Face. Distilled versions (7B, 14B parameters) run on consumer hardware and retain significant reasoning capability.
No. The model generates text tokens that represent reasoning traces — it's a learned behavior optimized by reinforcement learning. The appearance of deliberation is emergent from training, not evidence of inner experience. This is well-documented by Anthropic in their interpretability research.
Blogs
Comments are closed