Nestor G Pestelos Jr · ELI5

Phase 2: The User Interaction

Reasoning

How artificial intelligence moves from automatic word reflexes to deliberate, step-by-step logic and problem-solving.

01 / Fast Reflex vs Slow Logic

Easy questions use reflexes; hard problems need a napkin.

For simple questions, an AI uses fast pattern matching. For complex logic, it must write out its work step-by-step to avoid guessing.

FAST REFLEX (SYSTEM 1) "What is 2 + 2?" Answer: 4 (Instant recall) DELIBERATE LOGIC (SYSTEM 2) "What is 37 × 49?" Needs intermediate scratch work! If forced to answer 37 × 49 instantly without scratchpad steps, the model hallucinates.
Why can't models calculate silently in their heads?

A neural network gets a fixed amount of computing power per token. It cannot run a loop or pause to think internally. The only way it gets more computation time is by writing out its intermediate thinking steps into words.

02 / The Scratchpad

Reasoning models write their thinking before answering.

New reasoning models generate an internal "chain of thought" behind the scenes, testing formulas and verifying assumptions before showing you the result.

INTERNAL THINKING TRACE (HIDDEN) 1. Problem: 37 * 49 2. 49 = (50 - 1) 3. 37 * 50 = 1850 4. 1850 - 37 = 1813 ✓ Verified step-by-step USER OUTPUT 1,813 The user only sees the clean final answer, but the scratchpad made it possible.
What are thinking tokens?

In models like OpenAI o1 or DeepSeek-R1, the model emits hundreds or thousands of internal tokens before producing the final response. These thinking tokens are billed on compute, but are collapsed or hidden in the user interface.

03 / Self-Correction

Backtracking when an idea hits a dead end.

Reasoning models can catch their own mistakes mid-thought, say "Wait, that does not balance," and pivot down a better path.

Start "Wait, remainder doesn't fit." Proven Solution Found
How models learn self-correction

Models are trained using Reinforcement Learning with automated verifiers. If a math answer is wrong, the whole chain receives a zero score. Over millions of training rounds, the model discovers that actively questioning its previous steps leads to higher success scores.

04 / The Compute Tradeoff

Trading waiting time for rock-solid accuracy.

Older models answer in 1 second with a 60% chance of guessing wrong. Reasoning models think for 15 seconds to deliver near-perfect results.

STANDARD CHAT MODEL Speed: 0.8 seconds Math Accuracy: 62% REASONING MODEL (O1 / R1) Speed: 12.0 seconds Math Accuracy: 94% In artificial intelligence, patience directly purchases accuracy.
What is test-time compute scaling?

In traditional machine learning, models only got smarter if you made them bigger during training. Test-time compute scaling proved that a smaller model can outperform a giant model if you give it more time to think and verify its answers at inference time.