Deepseek is overthinking

>More importantly do you think if all those models worked 100% to specification there would be 0 basic hallucination errors? >Do you think that basic AI hallucinations, (the thing I am complaining about) has ever been a solved problem for any language model ever? While Large Language Models (LLMs) have shown significant improvement, their tendency to confidently hallucinate remains a challenge. This issue is multifaceted: "I don't know" is difficult to teach. Training LLMs on examples of "I don't know" as a valid response backfires. They learn to overuse this answer, even when they could provide a correct response, simply because it becomes a frequently observed pattern in the training data. LLMs lack robust metacognition. Current architectures struggle to facilitate self-evaluation. While reinforcement learning with extensive datasets holds potential for teaching LLMs to assess their own certainty, the necessary techniques and data are currently insufficient. Internal consistency remains a hurdle. LLMs are trained on massive datasets containing contradictory information (e.g., flat-earth theories alongside established science). This creates conflicting "truths" within the model, making its output context-dependent and prone to inconsistency. Training on fiction further exacerbates this "noise" by incorporating fictional world models. While improvements have been made by prioritizing data quality over quantity, this remains an active area of research. That being said, I tested the original numbers comparison on multiple locally hosted models on my own pc, and did not encounter a single wrong answer. All models responded that 9.9 is larger than 9.11. These were all small models wit 8B or less parameters. The smallest model I tested was 3B parameter starcoder2 with Q4K_M quantization, and even it got the answer right, despite being a very small model and relatively old on the scale of LLMs. I would not rule out user error or faulty quantization in cases where people encounter this error, especially when top-tier models like Llama 405B are considered.