I loaded 7 LLMs at once to test a simple coding task—most failed miserably

Here’s a reality check on today’s AI assistants: they’re still surprisingly brittle at the kind of tedious, small-bore coding tasks that would actually make them useful. I ran a deliberately simple test—asking multiple LLMs to translate plain English into a single line of Python API code for managing a calendar. The API had just two functions: add and remove an event. Simple, right?

The system prompt was straightforward: respond to every request with executable code. Yet the models consistently hallucinated non-existent API functions, ignored explicit instructions about the default 30-minute duration, or produced syntactically broken output. One model invented a create_event function out of thin air. When I pointed out the mistake with a short English sentence, the fix was often just as broken—sometimes introducing a new error while correcting the old one.

This wasn’t about solving Kaggle grandmaster problems. It was about dispatching drudgery. The kind of micro-task where you’d say “add lunch tomorrow at noon” and trust the machine to get it right. The results were a stark reminder that the gap between an impressive demo and a reliable tool remains wide. As one frustrated researcher put it, we risk building “artificial stupidity” when the model’s lack of formal logic collides with its confident tone.

I built a custom Gradio arena on a TPU v5e 2x4 with 128GB of memory to run this test, loading seven sub-10B parameter models simultaneously using Keras and JAX. The technical setup was solid. The models, however, were not. It suggests that for coding assistants to move beyond parlor tricks, the next leap needs to be in instruction-following reliability, not just benchmark scores.