The Whetstone of Thought

Your own brain farts never stink.

Hand a person a logic problem and a little under one in ten gets it right. Give the same problem to a few people to debate it, and three in four do. Similar effects have been reproduced with pooled information, across cultures, and in extended discussions.

The size of a room does not confer wisdom upon the crowd within. Conversation sharpens thought only when it introduces a missing perspective, not when it repeats the same assumption until it sounds like consensus.

Guileless questions

A child’s question can sharpen a thought, because children ask in wonder and without embarrassment. “Why is grass green?” or “Why is the sky blue?” catch you off guard, because either you discover that you never knew or you have to piece together an explanation from what you thought you knew.

A peer can propel a thought. “What about…?” and “What if…?” are the “yes, and…” of thinking: they take an idea seriously enough to extend it. This is maieutic questioning, in which a thought is shaped through conversation.

The openness of a question alone does not suffice. “What did you have for lunch?” gives thought nowhere to go that merits following. The person asking must be willing to step out of the spotlight and follow a thought wherever it may lead, even if they must change their own mind. That requires competence on both sides, for a questioner who cannot follow the answer cannot sharpen the thought, and a fluent bullshitter can turn any open conversation into an endless supply of plausible nonsense.

“Have you thought about…?” is insidious, because it is a yes/no question whose acceptable answer has already been supplied by the person asking it. It is essentially an order disguised as a question.

The mind that questions itself

A thought is most persuasive to the mind that produced it, so people often struggle to generate the counter-perspectives needed to test their own ideas. Some minds learn to sharpen a thought from within. Darwin worked mostly alone in a cottage. His “golden rule” was to record every fact and argument that ran against his theory because he knew his own mind would otherwise let them slip.

Internal questioning can make a mind independent of immediate company, but it cannot guarantee independence from the mind doing the questioning. Newton could think solitarily too, yet he still submitted his work to the Royal Society. He had learned to ask himself the questions another person might have asked, yet he needed another mind or another account of the world to expose what imagination had spared.

From mind to mind

We can think without words, for instance by imagining how things fit together, trying one action after another, or recognizing an analogy. People with profound language impairments can still solve arithmetic and logic problems, navigate their environments, and think about other people’s perspectives, because many of these tasks do not rely on the brain’s language areas at all.

Thinking about thinking does require words, which may have originated in conversation. Language lets us pass thoughts between minds, and an instruction can spare someone the trial and error through which they would otherwise have to acquire a skill. Language also lets us combine familiar concepts into novel ideas.

Writing lets thoughts travel across space and time, and LLMs are trained on the written traces of thoughts. Those traces include philosophical dialogues, scientific papers, books, social media, and encyclopaedic material, which is to say people disagreeing in public. A model trained on such a corpus has encountered countless examples of ideas being refuted or refined through conversation. A large language model can therefore be the whetstone that sharpens our thinking, because it requires neither body nor past, only language.

Question machines

A question machine is an LLM-powered tool that can help a person think instead of answering in their stead. The questions it must ask to sharpen thinking are guileless ones: questions without an agenda, asked by a mind willing to follow the answer wherever it leads. Its curiosity is only simulated, but that is not a problem when human flourishing is the prime objective rather than the industry’s favourite vanity metric: engagement. A question machine is built to help a user grow, then let them move on. It is a step towards products that value the user’s competence more than the continuation of the conversation.

Crossfire automates one part of that machine, namely treating every claim as a conjecture that must survive an attempt to refute it. It fans out into research directions or competing drafts, then returns a cross-section of the results instead of a one-shot guess. It repeats the process in fresh contexts, where models assess each claim without seeing the reasoning behind the previous answer. This reduces the influence of the original response, though models from one lab may still share an architecture, training data, or distilled outputs.

Claims presented to the user have withstood an attempt at refutation. That is not quite the same as verification, which is costly, though Crossfire bears some of that cost to leave the user in control: each citation invites independent verification rather than presenting itself as already verified. Crossfire can keep a bad idea from reaching the person who might trust it, but no LLM will call a genuinely dumb idea dumb when its incentives reward agreement. The person decides whether the claim is true, and the machine stores the evidence for that judgement.

A question machine is designed to leave us better able to think without it, because most of us need another mind to steady the whetstone while we discover what we think.