[Part 2] What is Deep Research? And why does 'delegating everything' fail?

1. Deep Research is a feature that combines 'smart search & summarization + trial execution'

1-1. Differences from standard chat AI

Comparison Item Standard ChatGPT ChatGPT with Deep Research Knowledge Acquisition Training data (up to around summer 2025) Internet search & summarization at the moment of being asked Code Execution None/Limited Python runs on the spot Flow to Output Answers only from 'inside its head' <Search>→<Summarization>→<Code Trial>→Answer
In short,
Deep Research is an 'AI version: an assistant that experiments on a laptop while researching'.

1-2. What is the benefit for data analysis?

1. Latest library trends are picked up immediately

2. Automatically cites 'techniques that were effective in similar competitions'

3. Runs sample code on the spot to show graphs and numbers

Even for Python beginners, since you can get a 'working template' immediately, the hurdle for 'copying code → running it' is significantly lowered.

2. Data competitions are a game of overcoming 3 mountains

1. Mountain 1: Understanding the data (EDA)

2. Mountain 2: Creating a baseline model

3. Mountain 3: Polishing the score (Tuning & Ensemble)

Tip
Just by telling the AI 'which mountain are you climbing now?', the suggestions you get back become much more accurate.

3. Why does 'leaving it all to AI' fail? 5 common failure patterns

# Common Failure What happens? One-point countermeasure 1 Vague goal Output becomes fuzzy State numerical goals first 2 Not providing data AI speaks based on imagination Pass at least column names & sample rows 3 Making it do everything at once Token overflow/cuts off midway Break down the process 4 Not sharing the score Improvement loop stops Paste LB results 5 Long-term automatic operation Stops due to errors in the middle of the night Check every 30 minutes

4. Success pattern: 4 steps of 'break down & back-and-forth'

Step 0: Decide goals and constraints and take notes (Human)
Step 1: Have it look at the data roughly (AI)
Step 2: Plan the battle strategy together (Human + AI)
Step 3: Write and run code (AI) → Return the score (Human)
↑───────Loop this 2-3 times────────↑
The point is to limit what humans do to only 'deciding and evaluating'.
With this, even at a level where you can write a little Python, you can sufficiently manage it.

5. Summary so far

• Deep Research is a super assistant that combines 'search, summarize, and experiment' into one set

• Competitions consist of three stages: 'understand, build, and refine'

• Delegating everything is a no-go, but breaking tasks down works

• By simply repeating the process of providing numerical goals and a portion of the data, receiving results, and getting improvement suggestions, a mid-tier score is within reach