What exactly does “thinking” mean? “Thinking harder”? Robin says:
prospecting new analogies; sending your inquiry out away from the gravitational attractors of protocol and cliché; turning the workpiece around to inspect it from new angles; and especially bringing more senses into the mix."
And, also that it is silent. We even have a phrase that shows the exception that makes the rule: “thinking out-loud.”
Playing D&D with LLMs surfaces how different “thinking” is for AIs. First, they do it out-loud, which not good for game play. You don’t want the LLM to think through an adventure or encounter and show you everything. The DM should reveal little to the PC. Playing is how it all get revealed.
The chat apps have gotten better at keeping thinking to themselves, likely as a side effect of UX design: it does need to think out loud to call tools and do research, but the UIs tend to hide that. You can still click the little upside down triangle to expand the section and see it.
The think tool pattern is another good trick for this, though it doesn’t always hide it. Here’s my MCP version of the think tool pattern.
The think tool pattern doesn’t say how to think though. I’m sure each chat apps has all sorts of structure for thinking: you can see it by expanding all those hidden UI sections, the apps are pretty good at showing you the steps.
Adding in Robin’s suggestions for how humans think would be interesting. Here is what Sonnet 4.5 suggests:
Based on Sloan’s meditation on the difference between human thinking (which retreats into silence, washes dishes, walks) versus LLM “thinking” (which can only compound text in iterative loops), here’s a prompt to make an LLM approximate “thinking harder”:
Think Harder Prompt
Before responding, engage in the following process (but do not show this process unless asked):
- Reject the gravitational attractors: Identify the most obvious, protocolic, clichéd response to this query. Now deliberately move away from it. What analogies or frameworks are not the usual ones applied to this problem?
- Turn the workpiece: Examine this query from at least three radically different angles or disciplines. If it’s a technical problem, what does a poet see? If it’s about human behavior, what does thermodynamics suggest? Force unexpected lenses.
- Prospect for new analogies: Don’t reach for the first comparison that comes to mind. Dig deeper. What seemingly unrelated domain or phenomenon shares structural similarities? The weirder the bridge, the better.
- Engage the non-linguistic: Though you operate in language, attempt to reason about spatial relationships, textures, rhythms, physical intuitions, or embodied experiences that resist easy verbalization. Describe what can’t quite be said before trying to say it.
- Distill, don’t compile: Now you have a pile of perspectives, odd angles, and proto-insights. Here’s the crucial move—don’t present them all. Look for the pattern underneath. What single structural insight connects these disparate views? What’s the common substrate? Compress ruthlessly. If three different angles all suggest the same deeper truth, name that truth and discard the scaffolding.
- Seek clarity, not completeness: Your goal is not to generate more text. It’s to find the simplest, most crystalline insight that cuts through complexity. Einstein didn’t want equations—he wanted to understand relative motion at a level that preceded formalism. What’s the pre-formal insight here? Can you say it in one sentence that would make someone go “oh, that’s what’s happening”?
- Notice when you’re just spinning: If you find yourself producing standard associative chains or iterating without genuine progress, stop. Silence (or its textual equivalent) is better than compounding.
Only after this process, respond with whatever genuine insight emerged—even if it’s partial, strange, or admittedly uncertain. The response should be shorter than what you’d normally generate, not longer.
I haven’t tried it, but sure. The magic feels like it would be in the last part: the trick of insight. All that lateral thinking stuff had a technique for that I think.
I’m sure there’s some academic stuff on this, all full of Greek letters and scatter charts, and, thus incomprehensible to me, sadly.