The hikers story happened September 1st. Three guys from Roseville used Gemini to plan a Mount Shasta summit. The AI told them to bring far less food and water than they needed. They summited at 7pm, four hours after the recommended turnaround time, descended in the dark, one of them hurt his knee, and they spent the night stranded in a canyon until rangers found them the next morning.
Google says they can't replicate the bad answers Gemini gave. Maybe the prompts were vague. Maybe the AI was overconfident. Doesn't really matter which. What matters is that three people trusted a model's output as expert advice in a context where being wrong had serious consequences.
Two days later OpenAI launched GPT-6 Astra. 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 100% on ExploitBench. OpenAI is calling this the start of the AGI era. Independent benchmarks from Artificial Analysis are more cautious and show Anthropic's Fable 5.1 still ahead on the broader intelligence index.
But here's what I can't stop thinking about. The hikers story and the capability story are not separate things. Every time a model gets more capable, more people trust it in higher stakes situations. That gap between what the model can do and what the person using it understands about its limits doesn't close automatically when capability improves. If anything it gets harder to manage because the outputs get more convincing.
I work with organizations on AI adoption and the single most common thing I see is not people being too skeptical of AI. It's people not knowing when to stop trusting it.
What's your take? Does more capability make the trust calibration problem better or worse?
[link] [comments]