How to Work Even in Your Dreams — How Humans Should Survive in the AI Era
The Combination Test Hypothesis — On What Sleep Does Where Coherence Loosens
Answers Come from the Pillow
During the period of working through my "What Einstein Missed" series one piece at a time, what unstuck the writing was not the desk but the pillow.
The questions I was turning over in that series were these. That gravity is velocity. The possibility of negative mass. Methods for detecting negative mass. There were often places where the answer was not visible at the desk.
For an hour or two before sleep, I turned those unresolved points over only in my head. Then the next morning, the moment I woke, a good idea was often already there. Not every time, but often enough.
What I had turned over before sleep had shifted into something close to an answer by the time I woke, and this happened often.
This is not coincidence.
The method that follows will be useful to workaholics who want to work even while they sleep.
Where Two Essays Meet
This essay starts at the intersection of two earlier ones.
The first is the Combination Test Hypothesis laid out in How to Manage Your Dreams. Dreams are an intensity-weighted random combination operation the brain runs each night, and the default is that the results evaporate. The fourth axis of that essay — pre-sleep emotion management — was a subtractive instruction: "Don't put bad input into the pool."
The second is the WWW theory presented in How Humans Should Survive in the AI Era. In an age where AI reads the already-drawn map remarkably well, the place humans hold is to accumulate many experiences because we don't know what work will come. And that work is decided by what one does every day.
This essay adds one line to those two declarations. Every day includes sleep. It pairs an additive instruction to what was a subtractive fourth axis — actively put the materials you need into the pool before sleep.
Actively Designing the Input
The combination test draws materials by intensity-weighted random sampling. Intensity is determined by recency, multiple pathways (seen / heard / written), repetition, emotion, and surprise. If you consciously turn over a problem you want to solve for an hour or two before sleep, the materials of that problem fill both the recency score and the active-processing score at once. They become priority candidates in that night's combination test.
The "hour or two before sleep" timing is decisive. About 50% of newly input information is forgotten within an hour, about 70% within 24 hours. While awake, that forgetting curve operates over the limits of working memory — four to seven chunks at a time. But when you cross into sleep, the curve is replaced by a different mechanism. Instead of the wear of the waking state, intensity-weighted sampling within sleep takes over. That is, material that entered the pool before sleep, instead of failing to last the next hour and disappearing, survives as a candidate for that night's combination test. The hour or two before sleep is the last window to fill the pool before forgetting halts.
This is not the mysticism of "sleep creates the answer." It is that I decide what materials enter the pool. Once the pool is set, the rest is the domain of probability.
The evidence is unexpectedly thick. In Wagner et al.'s 2004 Number Reduction Task experiment in Nature, the sleep group discovered the hidden rule more than twice as often as the awake group (60% vs. 22%). Cai et al.'s 2009 study in PNAS showed that REM sleep significantly raises performance on the Remote Associates Test — but the effect appeared only for problems with prior exposure. Meaning the effect is restricted to materials that entered the pool before sleep. Lacaux et al.'s 2021 study in Science Advances was more direct — people who spent at least 15 seconds in N1 right after sleep onset and were then woken showed nearly triple the success rate on creative problem solving (83% vs. 30%). Edison's technique of holding a metal ball and waking when it dropped was experimentally validated. Stickgold's 2000 Tetris experiment in Science is the most direct of all — strongly exposed pre-sleep stimuli appear as visual material at sleep onset (hypnagogic imagery), confirmed even in amnesic patients. That pre-sleep input enters the pool directly that night is a measured fact.
Historical anecdotes are plentiful. Kekulé's benzene ring, Otto Loewi's neurotransmission experiment (which won him the Nobel Prize), Mendeleev's periodic table, Paul McCartney's "Yesterday" melody. There is no need to believe every anecdote literally, but the pattern is consistent. They share one thing — they were people who had lived for a long time with the problem they wanted to solve laid out in their heads.
The Degree of Narrative Coherence Pressure
Here a comparison with AI becomes possible. And this comparison is where this essay meets the WWW theory.
Three positions, laid out:
First, AI is trained on coherent text corpora. RLHF reinforces that coherence further. Output is almost without exception trapped inside narrative coherence. If you ask an AI to produce output where "the dead and the living sit at the same table, and a room in a building you've never been to is the room from your childhood home," the AI tries one way or another to make sense of it. Coherence is the default.
Second, human waking thought also operates strongly under narrative habit. When we experience anything, we automatically frame it as "I saw, I acted, something happened." This automation itself is the default of waking consciousness.
Third, however, while a human is asleep, that narrative habit operates in a weakened state. It does not turn off entirely, but it does not operate at full strength either. The 2025 dream splicing follow-up by Stickgold and Sangodeyi in SLEEP Advances showed the size of that weakening — when dream reports with scene and plot discontinuities were cut and spliced with fragments from other people's dreams, judges identified intact from spliced at only 57%, near chance level. At the discontinuity points of dreams, coherence is essentially not traceable.
The third position is the core of this essay. To the extent that narrative coherence pressure weakens, room opens for combinations that did not surrender to coherence to be output. AI has almost none of that room; human dreams have plenty of it. This is not an abstract creative-superiority claim like "more creative than AI" — it is a difference at the level of architecture.
The consequence of this difference is this. AI interpolates within the training distribution. Every output has passed coherence checks within statistical patterns. AI almost cannot produce combinations not in the data, and even when it does, it immediately makes them make sense. Human dreams, by contrast, combine materials from one's own experience under weakened coherence checks, in ways never combined before. Combinations no human has ever placed side by side. That is how something not yet in the world becomes possible to output.
Where Randomness Is an Asset, Where It Is an Enemy
This technique works on one kind of work. The fact that the output is random determines that placement.
It is weak for exam study. An exam is the work of stably retrieving a fixed answer, and what stability requires is the memory consolidation of NREM sleep. This is a different mechanism that runs automatically over the day's learning, whether you turn over the problem before sleep or not. Pre-sleep study contributes a little to this mechanism, but the essence is daytime active learning, not the processing during sleep.
It is strong for creative work. The essence of creativity is finding unexpected connections between distant concepts. Remote association as Mednick defined it in 1962 is precisely that, and what Cai's study cited above measured is also that. A design problem that has not been cracking, the next paragraph of stuck writing, a research direction with no answer in sight — these are the right places.
When awake, the brain stays at near associations. This is a Bayesian-rational default — the closer the association, the higher the probability of being right. For everyday decision-making, this default is efficient. But for creative work, the same default becomes an enemy. Newness does not come from near associations. It comes from forcibly placing two distant concepts together. Sleep forcibly breaks that default. Intensity-weighted randomness places near and far associations as candidates with equal probability. Combinations that waking consciousness cannot reach are attempted within sleep.
To summarize: Exam study needs stability, so randomness is its enemy; creative work needs newness, so randomness is its friend.
The Posture of Laying Out
When turning over a problem before sleep, do not cling to solving it. Since answers are combined randomly within sleep, holding on to obtain will not produce a solution. Instead, lay out the structure of the problem in your head, set out the possible angles, and pass it to sleep without reaching an answer. The combination test works best in this posture.
The first axis of the original essay — emotional tag management — applies directly here. Material with a negative tag enters the pool with that tag and is sampled with negative affect. If the pressure to solve turns into frustration, that itself becomes input pollution. The very posture of setting it down without solving it is the best input.
Cognitive psychology research on the incubation effect arrived at the same conclusion. The consistent result from Sio and Ormerod's 2009 meta-analysis (117 studies) in Psychological Bulletin was that the probability of insight rises significantly when one returns to a problem after setting it aside.
Protocol
For an hour or two before sleep, consciously lay out the problem you want to solve in your head. Do not try to obtain the answer. Set out the possible angles, clearly recognize the point at which you got stuck, and pass it to sleep as is.
And — this part is decisive — keep paper and a pen next to the pillow. Write down whatever first comes to mind the moment you wake. A dream, a faint image, a sudden clear answer — write it down without exception.
If the last step is missing, the whole system does not work. The default for dreams is to evaporate. Exactly as the original essay laid out as hypothesis — the results of the combination test do not need to be kept, so they disappear. Even during the time they are remembered, they fade at 50% per hour. Stickgold's one line is exact.
"Dream recall improves not because dreams become more vivid, but because the brain learns that this material is worth preserving."
If the original is preserved, you can review it during the day. Most of it is useless combination. But one line becomes a clue that points toward the answer.
Conclusion
For an hour or two before sleep, lay out the problem you want to solve in your head.
Do not try to obtain the answer.
Pass it to sleep.
Write down whatever comes up the moment you wake, without exception.
That is all. It works for creative work, not exam study. The mechanism sits where two things meet — actively designing the input of the combination test, and the outside-of-coherence in human dreams that AI cannot reach. These two places operate simultaneously within the same person.
Recommended for workaholics who want to work even while they sleep.
Seungwon An / Wonbrand / https://wonbrand.co.kr