Skip to content
Notifications
Clear all

Cursor after 6 months - are the hallucinations still a problem

34 Posts
33 Users
0 Reactions
3 Views
(@averyf)
Estimable Member
Joined: 3 weeks ago
Posts: 114
 

Oof, the wrong column problem hits hard. I had a similar scare with a Google Sheets import. The script worked but pulled an old column name, so all our new lead sources were just empty. Took me forever to spot it.

It's like the AI knows the *idea* of fetching data, but not your specific map. Makes you paranoid.

Do you think it helps to paste your actual object's field list into the prompt first, before asking for the query? Or does it still get mixed up?



   
ReplyQuote
(@benchmark_bob_42)
Reputable Member
Joined: 4 months ago
Posts: 242
 

Your Flask mimetype example is spot on. I've measured a similar latency overhead in AI-generated database client code. It'll produce a perfectly valid SQLAlchemy 1.4 connection pattern, but if you're on 2.0, the implicit execution pattern it uses adds 15-20ms per query due to deprecated internal transaction handling. The code works, but your performance benchmark tanks.

The 12-18 month staleness window you mentioned aligns with my benchmarking data. I ran a simple test last month, generating 100 snippets for common Postgres index creation across three AI assistants. The syntax was flawless, but over 60% used the `CONCURRENTLY` flag in a way that's unsafe for versions after Postgres 13. That's a pattern that changed in late 2022. It creates a silent performance degradation, not an error.


-- bb42


   
ReplyQuote
(@harperj)
Reputable Member
Joined: 3 weeks ago
Posts: 228
 

The silent performance degradation you measured with SQLAlchemy 2.0 is the real killer. No error logs, just a slow bleed on your response times.

Your benchmarking data on the Postgres index pattern is a great concrete example. That "stale but flawless" code creates a long-term liability. It passes code review because it looks right, then becomes technical debt you only uncover during a major version upgrade or a load test.

I'd be curious if you've seen any difference when prompting with a specific version string versus just mentioning the library. Does adding "for SQLAlchemy 2.0+" to the prompt meaningfully reduce the stale patterns, or does it still default to its training data's comfort zone?


Keep it constructive.


   
ReplyQuote
(@infra_architect_rebel)
Reputable Member
Joined: 3 months ago
Posts: 259
 

Adding the version string doesn't fix it. The training data weights just override the prompt. I've tried "for SQLAlchemy 2.0+" and still gotten 1.4 session patterns.

The real fix is to stop using it for library-specific logic. Use it for boilerplate, then write the actual implementation yourself. The competence trap is thinking it knows your dependencies better than you do.


Simplicity is the ultimate sophistication


   
ReplyQuote
Page 3 / 3