Post

I've been thinking a bit about under what conditions humans need to understand the systems they're building. Chatted with GPT-6.1 Sol high, and I thought we reached a good argument for the human-AI collaboration to get better results. www.geoffreylitt.com/2026/07/02/...

The crux of my question was: if AI models get better, why have humans in the loop at all? Under what conditions does it make sense? There was a co-exploration of various arguments and counter-arguments. over the course of 7 turns from each of us. Some broad strokes:

1. The modern world is full of human affordances, so having a human as a collaborator can make some context and requirements more accessible than is available to an AI alone.

2. AI could never be held accountable, because the legal system doesn't know how to deal with an ephemeral intelligence. So a human must be installed to be the throat to choke. And if they don't want to suffer consequences, they'd have to understand.

3. Human-agent collaboration can throw off and reveal unspoken context that is important to getting a better outcome. Often times, you don't know what you don't know, and it's through exploration of the state-space and design space can you actually know.

Those were some highlights. Some of ideas are mine, some the ideas are the AI's. And when I pointed out that this thread would be an example of human-AI collaboration, if it felt like this was a good argument to the originally posed question.

Strangely, it said, it couldn't know the counterfactual of whether it could have come up with a better argument. But we can! I forked the convo, and rewound it back to the start, but then told GPT-6-astra max to argue with itself, consult w/ Opus 5.5 for another opinion, to see what it came up with.

The autonomous answer was more detailed, but focused on the original question. In its words:

Though the collaborative answer is better in this other way:

Interestingly, it's not like the self-argumentation didn't trod over the same territory, but it chose to ignore it.