Metacurity

"They often generate code that fails to perform as intended, but which on the surface seems to run successfully, avoiding syntax errors or obvious crashes."

AI Coding Assistants Are Getting Worse
https://spectrum.ieee.org/ai-coding-degrades?utm_source=homepage&utm_medium=hero&utm_campaign=hero-2026-01-10&utm_content=hero1

AI Coding Degrades: Silent Failures Emerge

One AI coding assistant power user says the tools are hitting a plateau, and some are even declining. What's causing this unexpected twist in tech?

IEEE Spectrum
Jan 10 at 12:48pmWeb
Show thread~w00p~Jan 10
@metacurity I had a similar experience. Played around with llms for my Bachelor's thesis and ended up dos'ing myself (😂). Syntax was correct but there was a logical mistake in the returned query.
Everything that comes out of an llm must be checked and double-checked.
Show thread~w00p~Jan 10
@metacurity I'm pretty certain that in the future there will be spectacular failures (stuff like this https://www.simscale.com/blog/nasa-mars-climate-orbiter-metric/) related to llm generated code.
It's a bit scary even
How NASA Lost Its Mars Climate Orbiter From a Metric Error

The Mars Climate Orbiter was launched by NASA in 1998 to study the Martian atmosphere and surface changes but was lost due to a metric error.