I stayed up past two in the morning last night watching a machine change its mind, because DeepSeek prints its entire inner monologue before it answers and I had asked it something about distributed consensus, a topic I knew nothing about, and before it ever arrived at an answer I got a front row seat inside its head. It tried an approach, hit a contradiction, wrote wait, that cannot be right, and backed up, it enumerated three possible designs and killed two of them for reasons it explained mid sentence, it second guessed its own arithmetic and redid the calculation, and somewhere around minute four I realized I was not using software anymore, I was being taught, because in twenty minutes of reading its doubts I understood more about how to attack an unfamiliar systems problem than years of consuming polished answers had given me. That experience is available right now to anyone with a browser, and it is exactly the experience the biggest AI companies in America have decided you should never have.

Here is the split that nobody talks about loudly enough, the American frontier models think in private and hand you a press release about their thoughts, while the Chinese models print the whole messy monologue and let you follow along. OpenAI said it themselves in September 2024, in a section titled Hiding the Chains of Thought, where they announced they would not show the raw reasoning of o1 and would display a model generated summary instead, and the sentence deserves to be read slowly because of the three factors they weighed, user experience, competitive advantage, and chain of thought monitoring, the middle one is doing all the quiet work. They even admit the raw chain of thought lets them read the mind of the model, which tells you the thoughts exist and are legible, they have simply reserved mind reading as a shareholder benefit. Google ships thought summaries, which is the same product with friendlier branding, and meanwhile DeepSeek publishes complete traces alongside open weights you can download, and Qwen’s QwQ reasons in public the same way, so the entire question of whether a reasoning model should share its reasoning has quietly become a geography question.

There is an old saying about fishing that you have definitely heard, give a man a fish and you feed him for a day, teach a man to fish and you feed him for a lifetime, and almost everyone repeats it as an ancient Chinese proverb attributed to Lao Tzu, with Maimonides and the Bible occasionally drafted in for variety. The trail leads somewhere much funnier, because the earliest known version comes from Anne Isabella Thackeray Ritchie, daughter of the author of Vanity Fair, writing in her 1885 novel Mrs. Dymond, if you give a man a fish he is hungry again in an hour, if you teach him to catch a fish you do him a good turn. Researchers who chased the Chinese attribution found no ancient source at all, and the standard Chinese phrasing may well have traveled east from the English, which means the proverb spent a century wearing a false passport, and there is something almost poetic about the fact that today the labs actually teaching the world to fish are the Chinese ones while the American labs, whose founding myth involves openness, stand at the counter handing out fillets with the skin removed.

What reading a full trace does to a person like me is hard to overstate, because I am a normie in most domains and the trace is the first tool I have ever held that teaches the shape of thinking itself rather than the content of one field. Watch any good model reason about organic chemistry and you absorb how chemists attack problems, generate candidates, prune by first principles, sanity check against known values, and watch the same model reason about a legal clause and you absorb a different discipline built on definitions and precedent, same machine, different epistemology, handed over intact. This is precisely what I always wanted and never expected to get, which is to crawl inside an expert’s head and watch them work a problem from confusion to clarity, the way apprentices once stood behind master craftsmen for seven years hoping osmosis would do its thing, except now the apprenticeship costs electricity and covers every trade at once.

I was going to apologize for the digression I am about to make and then I remembered nobody reads apologies, which is that humans are memers down to the bone and always have been, we tell ourselves stories about originality while our jokes mutate through group chats and our ideas arrive through citations and our taste is a collage of everyone we have ever admired. Almost nothing any of us thinks is original, it is copied, remixed and recombined, and the copying is not a bug, it is the entire mechanism by which culture moves, because imitation performed deliberately has another name and that name is learning. So when a model shows me its thinking, it is doing the oldest human thing there is, demonstrating a skill slowly enough for the next monkey to steal it properly.

Which brings me to why the hiding bothers me more than the benchmarks ever could, because a lab that shows you only conclusions is training a planet to consume verdicts, and a generation raised on verdicts gradually loses the muscle for deliberation, the way nobody alive remembers how to sharpen a scythe or navigate by stars because machines took those jobs before anyone mourned them. OpenAI’s stated excuse includes protecting their work from distillation by competitors, and fine, I understand moats, but notice who the distillation protection actually hurts, because the rival labs will extract the reasoning capability regardless while the eight billion ordinary people who could have learned to think from these traces get locked out permanently, so the moat keeps out the students and admits the thieves. A handful of companies end up as the sole custodians of the intelligence frontier, renting cognition by the token to everyone else, and the rest of humanity slides gently into becoming an audience for its own replacement, nodding along at summaries of thoughts it will never be shown.

Ritchie’s line needs one update for our century, because giving a man an answer creates a customer who returns tomorrow hungry for another answer, and teaching him to think creates a rival who might one day out reason you, and I suspect the fear of that rival is the real product strategy hiding inside the word safety. The labs that conceal their chains of thought have decided we are customers, and customers pay rent on their own minds forever, while the ones that publish everything have decided we are students, and students are dangerous because they graduate, and I for one intend to graduate, one downloaded trace at a time.

Tell me the best thing you ever learned by watching someone think. I am @troysk704.