What Generative AI Can and Cannot Do · AI and Generative AI Introduction in Generative AI | Generative AI & LLMs

Strengths and Weaknesses

Good at	Weak at (without help)
Writing and rewriting text: emails, replies, product descriptions	Exact facts it never saw, or facts after its training cut-off
Summarizing long documents	Exact arithmetic on big numbers, counting letters
Translating (English, Hindi, Tamil, ...)	Knowing your private data (orders, policies) unless you send it
Extracting fields from messy text into JSON	Guaranteeing the same answer every time
Writing and explaining code	Taking real actions (refunds, bookings) unless you give it tools
Classifying tickets, reviews, intents	Saying "I do not know" when it should (it may hallucinate)

Use Code for Exact Work

Example

An LLM reads text as tokens (chunks like straw + berry), not as single letters, so questions like "how many r in strawberry" are surprisingly hard for it. Plain code does exact work perfectly. Good GenAI apps combine both: the model for language, code for exact steps.

word = "strawberry"
print("letters:", list(word))
print("count of 'r':", word.count("r"))

order_total = 1499 * 3 + 249          # 3 shirts + delivery
gst = round(order_total * 0.18, 2)
print("order total:", order_total, "GST 18%:", gst)

# A model might see the word as chunks like these (made-up split for illustration)
print("possible tokens:", ["str", "aw", "berry"])

Output

letters: ['s', 't', 'r', 'a', 'w', 'b', 'e', 'r', 'r', 'y']
count of 'r': 3
order total: 4746 GST 18%: 854.28
possible tokens: ['str', 'aw', 'berry']

Note: Later you will give the model tools (Python functions) such as a calculator or an order lookup, so it can ask your code for exact answers instead of guessing.

Exercise

Exercise: For each task, say whether you would use an LLM, plain code, or both: (a) compute a refund of 2 items at ₹799, (b) write a polite apology for a late delivery, (c) read a customer email and fill an order-ID field.

Show answer

(a) Plain code: exact arithmetic (2 * 799 = 1598).
(b) LLM: open-ended writing.
(c) Both: the LLM extracts the order ID from free text,
    code validates its format (e.g. SE-1234) and looks it up.