And they went ahead and released it anyway. News Update, Friend Apple's latest venture into AI, called Apple Intelligence, has been largely underwhelming. Its news summaries, in particular, have drawn significant backlash for mismanaging headlines and providing inaccurate information, prompting Apple to put the entire program on hold this week for necessary fixes. None of this is particularly surprising. Issues like AI "hallucinations" are a known challenge for all large language models, and it remains unsolved—if it can be solved at all. However, launching its AI model appears especially reckless given that Apple engineers had previously highlighted serious flaws in the technology. This caution was articulated in a study released last October. The still-unpublished research, which examined the mathematical "reasoning" capabilities of some leading industry LLMs, contributed to the understanding that these models do not genuinely reason. "Instead, " the researchers noted, "they seek to emulate the reasoning steps found in their training data. " Math Challenges To evaluate the AI models, researchers tasked them with solving thousands of math problems from the commonly used GSM8K benchmark dataset. A straightforward question might be: "James buys 5 packs of beef that weigh 4 pounds each. The price of beef is $5. 50 per pound.

How much did he spend?" Some questions were slightly more complex, yet still manageable for a reasonably educated middle schooler. The researchers highlighted gaps in the AI models with remarkable simplicity: they merely altered the numbers in the questions. This approach mitigates data contamination—meaning the AIs haven't encountered these specific problems before in their training data—while not increasing the difficulty of the problems. This adjustment alone caused a slight but significant drop in accuracy across all 20 tested LLMs. However, when the researchers escalated their method by also changing names and introducing irrelevant details—like specifying that a handful of fruits were "smaller than usual"—the decline in performance was, as the researchers put it, "catastrophic, " reaching as high as 65 percent. Performance varied among models, but even the most advanced, OpenAI's o1-preview, experienced a drop of 17. 5 percent, while its predecessor GPT-4o suffered a 32 percent decline. Learning from Patterns The implications are stark. "This exposes a crucial flaw in the models' capacity to identify pertinent information for problem-solving, likely because their reasoning isn't formally structured in a traditional sense, but primarily relies on pattern recognition, " the researchers asserted. In simpler terms, AI excels at seeming intelligent and often provides correct answers!However, once it is unable to replicate specific data, it falters significantly. You would think such findings would raise significant doubts about relying on an AI model to generate headlines—rearranging words without truly grasping how that alters the overall message—yet that doesn’t seem to be the case. Apple was aware of the critical issues that have persisted across every LLM and launched its model regardless. To be fair, this has become the standard practice across the AI industry. More on AI: Disturbing New Startup Deploys AI Agents to Flood Reddit with Promotional Posts for Clients' Products.