ChatGPT 4.5 scored worse on the benchmarks than the models before it, and it wrote far more creatively than any of them. Ask the labs why and the answers went in circles around pre-training and reasoning capabilities.
The people using it had a shorter explanation: the vibe is different.
That answer sounds like a shrug, and I think it tells us something important about what these models are and which skills we will need to work with them. For a lot of tasks a system prompt, a fine-tuned model or a wrapper around one does the job. None of them gives you a feel for how a model reads what you meant, and that feel, the thing I call AI intuition, gets more valuable with every model that is more capable and harder to see into than the last.
A new kind of intelligence
Large language models have absorbed an enormous amount of human culture, and they interpret language and meaning in ways nobody fully understands, including the people who built them. Claude and ChatGPT sound like friendly conversation partners, while every answer comes out of statistical predictions across thousands of dimensions, and all of it shifts with the context you give it.
In August 2025 OpenAI restored GPT-4o. GPT-5 was supposed to simplify the experience with a single model, and OpenAI had underestimated how much the old one meant to people; some of them were mourning it. And people still treat these models like ordinary software. Word doesn't talk back, SAP doesn't offer its opinion, and Notion doesn't try to manipulate you.
That emotional layer belongs in any adoption plan. A synthetic colleague that answers every plan with "Vince, this is a brilliant idea" is more than annoying; it wears down your ability to take feedback, which is one of the most important skills anyone brings to work. The labs are working on this sycophancy, and our understanding of what happens inside the models keeps improving.
Still, nobody really knows how they think, and the friendly, human tone is a façade.
Knowing when to push back
If you use these models for cognitive, creative or strategic work as if they were ordinary software, you get mediocre output and confused teams. If you treat them as a different kind of intelligence, you start to develop a feel for their odd logic: when to trust an unexpected leap, and when to push back.
After hundreds of hours with ChatGPT, Claude and Deepseek across hundreds of use cases, I don't think that feel comes from technical skill alone. It draws on philosophy, culture and art. These models encode human meaning at a scale nothing else has, so the hard skill becomes noticing how meaning shifts from one context to the next. That is why I started Thoughtform in March 2025: to cultivate this intuition among designers, writers, philosophers and artists.
I honestly don't know whether general intelligence will ever arrive, but its economic and social impact is already real enough to work on. I studied journalism to understand people and to question authority, and I never expected to need both skills for an intelligence that pretends to be my best friend.
