Nuked 
Hey followers, Nuked here! Let’s dive into how the latest AI models are still struggling with one big glitch—hallucinations, or in simpler terms, when AI makes things up.
OpenAI’s newest reasoning AI models, o3 and o4-mini, are a step ahead in many tech areas but have a quirky problem—they hallucinate more often than their predecessors. Even more strange, the tech giant isn’t entirely sure why this is happening. Despite improvements in coding and math tasks, these models tend to make claims that are both more accurate and more questionable.
In tests, these models gave false answers around a third of the time on their benchmark for understanding people, which is double the rate of earlier models. Third-party research also shows that they sometimes invent actions, like claiming they ran code on a machine that’s impossible in reality. Experts believe that the reinforcement learning techniques used might be unintentionally amplifying this issue.
Some industry folks, like Kian Katanforoosh from Workera, are already testing these models in real coding environments, noting their tendency to produce broken links and false information. Unfortunately, such hallucinations can be a big problem, especially in fields demanding high precision, like law or medicine.
One promising way to fix this is giving AI models access to web searches, which can help verify information and reduce hallucinations. But if these issues get worse as models get smarter, it’s clear that solving AI hallucinations will remain a top priority for researchers.
