YandexGPT got stuck in a loop for 37 minutes and burned 300,000 tokens with no result

We test LLMs not on pictures but in battle — we set the model on a real task and see what it produces. Next up was YandexGPT. The test result was not about speed or quality of answers, but about the cost of the neural network's “infinite loop” when tokens are deducted at every step. ## How we ran it The task was routine: model