Report finds newer inferential models hallucinate nearly half the time while experts warn of unresolved flaws, deliberate deception and a long road to human-level AI reliability
Agreed. There was a time when it worked impressively well, but it’s become increasingly lazy, forgetful, and confidently wrong, even missing obvious explicit prompts. If you’re using it thoughtfully as an augment, fine. But if you’re relying on it blindly, it’s risky.
That said, in my experience, Anthropic and OpenAI are still miles ahead. Perplexity had me hooked for a while, but its results have nosedived lately. I know they tune their own model while drawing from OpenAI and DeepSeek vs their own true model but still, whatever they’re doing could use some undoing.
Agreed. There was a time when it worked impressively well, but it’s become increasingly lazy, forgetful, and confidently wrong, even missing obvious explicit prompts. If you’re using it thoughtfully as an augment, fine. But if you’re relying on it blindly, it’s risky.
That said, in my experience, Anthropic and OpenAI are still miles ahead. Perplexity had me hooked for a while, but its results have nosedived lately. I know they tune their own model while drawing from OpenAI and DeepSeek vs their own true model but still, whatever they’re doing could use some undoing.