LLM vs SQLite

For anyone using AI tools, here is something for you to try.

Imagine you have SQLite database with a following schema & data:

create table addresses (ipv4 blob not null);
insert into addresses (ipv4) values (x'c0ae1337');

The task is pretty simple:

Return IP v4 address in a human readable format.

Answers:

select
  printf('%s.%s.%s.%s',
    '0x'||substr(hex(ipv4), 1, 2) ->>'$',
    '0x'||substr(hex(ipv4), 3, 2) ->>'$',
    '0x'||substr(hex(ipv4), 5, 2) ->>'$',
    '0x'||substr(hex(ipv4), 7, 2) ->>'$'
  ) as ipv4
from addresses;

How I ended up with IPv4 address stored as blob in SQLite?

Well, I was running some experiments and I didn’t want to deal with string manipulations for this task. After all IPv4 is just 4 bytes, right?

Later, when I wanted to retrieve the address in a human-readable format, it turned out to be a bit convoluted in SQLite. I could have handled this easily in PostgreSQL, but SQLite doesn’t have built-in support for IP types.

So, I decided to try some AI tools. After 15 minutes I gave up and ended up reading the documentation. Do I regret the wasted time? Yes – AI gave me dozens of queries, all of which returned incorrect results, contained syntax errors or were simply hallucinations.

Why am I writing this post?

I ran a little quiz with about 20 friends, asking them to solve the same task. Only one person got the correct answer (but they failed to generate a proper query with AI).

Most people quickly answered 0.0.0.7 or another wrong IP without validating their results. Some were just copy-pasting invalid queries. I was frustrated because how many of you blindly trust AI tools.

I am not saying that all those AI tools are bad, I just want you to use those tools wisely. LLMs were trained on open-source projects, but who graded the quality of those projects?

If you don’t want to think yourself, then please don’t name yourself a Software Engineer (because you are a Prompt Engineer).