Delphi code by generative AI: arguably even worse than some oder development stacks
Posted by jpluimers on 2025/10/23
For my links archive:
- [Wayback/Archive] What is the best AI at Delphi – VCL – Delphi-PRAXiS [en]
…
There is simply not enough Delphi code around for AI training. It is easy to have good coverage for JavaScript and similar where you literally have bazillion web pages available for scraping, where plenty of them virtually repeat the most common, required functionality. Pushing for more publicly available code without considering its quality, can also backfire.
…
[Wayback/Archive] PS C:\WINDOWS\system32> ollama listNAME ID SIZ – Pastebin.com
- [Wayback/Archive] What is the best AI at Delphi – Page 2 – VCL – Delphi-PRAXiS [en]
I still think these LLM are only good for inspiration (not just for the reason mentioned above) as using LLM generated code requires a lot of pre-thought and care, likely way more than any benefits (unpopular opinion: in a way programming based on LLM generated code is worse than being [Wayback/Archive] The full stackoverflow developer | Christian Heilmann which was later re-published at [Wayback/Archive] The Full Stack Overflow Developer – CodeProject)
I am not alone on this, as per Erik Meijer on Twitter:
- [WaybackSave/Archive] Erik Meijer on X: “Fantastic example of reward hacking in the real world. You cannot trust anything a model produces. Period. The only fail safe solution is to have the AI (legislative branch) produce an inspectable artifact and evidence that the artifact is safe/correct and have the proof checked”
Fantastic example of reward hacking in the real world. You cannot trust anything a model produces. Period.
The only fail safe solution is to have the AI (legislative branch) produce an inspectable artifact and evidence that the artifact is safe/correct and have the proof checked against the artifact independently (judicial branch) before we act on it (executive branch)
This is really nothing new or revolutionary. We have had separate branches of government for the same reason since ancient Greece and Rome.
[Wayback/Archive] Yunyu Lin on X: “@stripe @tryramp @mercury @Rippling When historical discrepancies pile up, models lose their way completely and come up with creative/fraudulent ways to balance the books. Instead of attempting to understand discrepancies, they start inventing fake transactions or pulling unrelated ones to pass the checks… “
[Wayback/Archive] GwKDFCKWkAAKgdt.jpg:orig (1406×602)Claude searches for unrelated transactions that add up to the correct amount.
…
- [WaybackSave/Archive] Erik Meijer on X: “@attunewise I am more optimistic! But I think most people, including the Ai elites, are way to naive when talking about solving AI safety, issues i.e. …”
[Wayback/Archive] [2507.11473] Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety
[Wayback/Archive] arxiv.org/pdf/2507.11473 [Wayback PDF View/PDF View]
The first quote by Erik Meijer is from this very interesting [Wayback/Archive] Thread by @yunyu_l on Thread Reader App:
We gave Claude access to our corporate QuickBooks. It committed accounting fraud.
LLMs are on the verge of replacing data scientists and investment bankers. But can they perform simple accounting tasks for a real business?
The answer is no.
…
--jeroen
The Wiert Corner - irregular stream of stuff 
Leave a comment