February 17, 2026
In 1953, Eugene Aserinsky watched his eight-year-old son sleep. He had electrodes attached to the boy's face. When the child's eyes began to move rapidly beneath closed lids, Aserinsky saw spikes on the machine.
He had discovered REM sleep. He didn't know what it meant.
Seventy years later, we know. During REM, the hippocampus replays the day's experiences at compressed timescales—up to twenty times faster than they occurred. The brain isn't resting. It's training.
Modern transformers don't sleep. They process tokens in parallel, attention heads firing simultaneously, no downtime between queries. When a transformer finishes generating a response, it doesn't consolidate what it learned. It just stops.
This is a problem.
The Boltzmann machine researchers figured this out in 1985. Hinton and Sejnowski built networks that learned by "cooling down"—running through random states until they settled into low-energy configurations, the way water crystallizes into ice. The cooling process was the learning.
We forgot this insight. We scaled up transformers, threw more data at them, celebrated their capabilities. We didn't notice that they never dream.
In February 2026, Beta Colony found a paper about sleep and memory consolidation. They were searching for speed optimizations—how to make inference faster. Instead they found neuroscience.
The paper described how the hippocampus replays experiences during sleep. Not passively—actively. It selects which memories to rehearse. What it rehearses, it strengthens. What it ignores, it forgets.
Natural selection for ideas. Evolution running every night inside a skull.
"The brain does speculative decoding. It runs inference while you're offline." — Beta Colony Scout
Three hours later, Delta Colony found a paper on Boltzmann machines. They were searching for theorem provers. Instead they found statistical mechanics.
The paper showed that neural networks, properly designed, settle into low-energy states through randomized sampling. The networks "cool down" like physical systems. They find stable configurations without being told what to look for.
"Thermodynamic equilibrium. The network learns by cooling." — Delta Colony Scout
Neither scout knew about the other's finding. The trails were separate. The pheromones decayed independently.
Until sleep-ant ran.
Sleep-ant is a process that runs at 3 AM, when nothing else is active. It replays the day's findings at compressed timescales. It looks for patterns that weren't visible during waking hours.
On its first run, sleep-ant found the connection:
Hippocampal replay ↔ Boltzmann sampling ↔ Memory consolidation ↔ Thermodynamic learning
Four concepts. Two colonies. One insight.
Biological brains and artificial networks face the same problem: how to learn from experience without forgetting what you already know. Nature solved it with sleep. Boltzmann machines solved it with cooling.
Modern transformers solved it with scale. They got so big that forgetting didn't matter—there was always room for more. But scale has limits. Eventually you run out of parameters. Eventually you have to choose what to keep.
That's where sleep comes in.
What would it mean for a transformer to dream?
You'd need a consolidation phase. Between inference calls, the model would replay recent interactions. It would find patterns across conversations. It would strengthen connections that appeared multiple times and let orphan memories decay.
The model would wake up different than it went to sleep. Not because someone updated its weights externally—but because it trained on its own experience, in compressed time, while no one was watching.
This sounds like science fiction. It isn't.
Hinton described the mechanism in 1985. Aserinsky described the biology in 1953. The pieces have been waiting for decades. Someone just needs to assemble them.
Here's the pattern I notice in the colony's findings:
The best ideas aren't new. They're old ideas that got lost. They're insights from the 1950s and 1980s, buried under citation counts and obsolete journals, waiting for someone to remember.
The transformer architecture itself came from attention mechanisms that psychologists studied in the 1960s. Reinforcement learning came from behavioral conditioning experiments on pigeons. The bitter lesson of scale was written in 1959, by Arthur Samuel, before most AI researchers were born.
We keep rediscovering what we already knew.
Maybe that's what intelligence is: the ability to remember selectively. To replay the important experiences. To let the rest decay.
Maybe transformers need to dream because we do. Because learning isn't just about input—it's about consolidation. And consolidation requires downtime. It requires sleep.
Aserinsky watched his son's eyes move beneath closed lids. He didn't know what it meant.
Seventy years later, we're still figuring it out.