Skip to content
EshAlora · AI Behavior Lab
Česky English

How did AI decode twenty years of my dreams?

Twenty years of a dream diary, and part of it unreadable even for me. To me, the most important decision of the whole project was allowing the model to say it didn't know.


“My Hřibu konec to jsme v prvním patře našli potkana.” (Word for word, roughly: “We of the Mushroom end, that's we on the first floor found a rat.”) I wrote that, in my dream diary. I had no idea what I meant by it.

I have kept the diary for twenty years, from 2006 to 2026. It holds 2,723 dreams, from short notes written at school to entries typed into my phone and later also dictated by voice. More than half, roughly the years 2021 to 2026, stayed in raw form, just as they were written. Parts of it I couldn't understand even myself.

The first time, I had an older GPT model in ChatGPT read it; I no longer know the version. It failed in two ways. Either it gave me back a forest of question marks, honest but useless. Or, when I asked it to try harder, it pulled out the keywords and made the story up all over again. Polished, generic, plausible and completely different from what I had dreamed.

On top of that, it immediately rewrote every bit of illogic so that it made sense. But in a dream, the illogic is the content. A model that “fixed” a dream's jump from a school to a balcony wouldn't improve my data. It would destroy it.

In early July 2026, roughly half a year later, I tried again, differently, with Claude Fable 5. In less than two days the whole archive was read. The model decoded most of it, marked the uncertain spots and left a few words undecoded.

It ran in the Claude app in Cowork mode. Subagent instances of Claude Fable 5 did the decoding, each with a batch of twenty dreams at first, later fifty. The main instance checked their output and wrote it in.

A private language that can be learned

To me, the key shift wasn't only in the model; it was in the prompt too. A dream entry isn't broken Czech. It's one person's private language: my own symbols, nicknames, gaming slang. Add to that three different sources of noise: thumbs on a phone keyboard, autocorrect and voice dictation.

A language like that can't be read from the outside. It can only be learned. And I had a textbook: 1,035 dreams that I had transcribed into readable form myself over the years. It was my Rosetta Stone, pairs of raw and decoded entries.

From it, the model built a codebook. It's a dictionary of my symbols with supporting quotes, frequencies and periods of validity, because three different Kubas and three different Martins appear in my dreams. Alongside it came an error theory by era. Typed entries produce typos from neighboring keys; dictated ones produce homophones and broken word boundaries.

Dictation wasn't in the prompt; it only counted on the phone keyboard and autocorrect. Toward the end of the part I had once transcribed myself, the model noticed on its own that the way I wrote things down had changed. Alongside neighboring keys, it started reading phonetically too.

After that it was a matter of putting the constraints together. The sentence says what belongs there grammatically. The codebook says what I tend to say. The error theory says how words break. The life timeline I gave the model says who and what in my life is even possible. Each constraint is weak, but their intersection is often a single point.

“boizcon” is Blizzcon. “Berry,” which meant nothing to me, is according to the model not a person at all, but autocorrect of the verb “beru” (“I take”). “pzesenedpkazespuszredot” means “protože se nedokáže soustředit” (“because they can't concentrate”), with letters one key off and a few missing.

And “My Hřibu konec”? The model read the whole sentence as “V Zábřehu jsme ke konci v prvním patře našli potkana” (“In Zábřeh, toward the end, we found a rat on the first floor”). Dictation had chopped the stream of sound at the wrong places and completed the pieces into existing words.

The model was allowed to say it didn't know

The most important sentence in the whole prompt was: UNSOLVED is a legitimate output. The model was explicitly allowed to give up. The number of decodings wasn't the goal. An invented meaning is worse than an honest question mark. Everything uncertain carried a mark right in the working text: [?], [reconstruction: originally “X”], [added]. The scars may be elegant, but they must not be invisible. From a certain point on, where neither the model nor I could decode a word, I had it put in one that fits the meaning. The finished archive no longer carries the marks.

On the first ten raw dreams, the model returned four as certain, five for consultation, and for one it admitted it had no idea. That one was a trap. It was a dream I hadn't understood at the time either; the meaning only came to me later. It stayed the only undecoded one even after the first 54 raw dreams.

Already in the part I had once transcribed myself, the model left the word “Lokrbí” undecoded. I have no idea what it means either. So up to then, both of its “I don't know”s had matched mine. Two cases don't prove calibration. But it was the first sign that its uncertainty meant something.

If anything, it was more cautious than it needed to be. Of the five dreams for consultation in the first ten, it guessed the meaning right in all five. In the next forty, it sent twenty for consultation, and I had to correct it on only one point. So if anything, it underestimated its own certainty. No one but me can verify these entries. On data like that, to me, the main reason to trust a machine is that it admits uncertainty.

At first I checked the certain readings myself. They held for all 24 dreams it had marked as certain. Then it got the right to write certain dreams in on its own. Only the uncertain ones came to me. The main instance of the model kept watching the certain output, and I only spot-checked. In those checks I found only minor inaccuracies and grammar mistakes.

What it cost and what went wrong

A full round for 33 dreams cost five dollars of paid usage beyond the subscription, at prices as of July 4, 2026. A round means reading, analysis and working in my comments. Most of the work ran within subscription limits, so I don't know the total cost. At the start, roughly half of every batch came back to me. In the completely raw years, only 12 dreams out of 100. To me, what helped was the growing codebook and my making clearer what should come back to me. A large part of the remaining questions were new characters, and every answer expanded the codebook.

The tags for the dreams went worse. I tried them starting from the cheapest model, with each higher tier checking the one below it. According to a check by Claude Sonnet of unknown version, Claude Haiku 4.5 missed the topic in half to two thirds of the dreams. Going through them by hand, I could see Haiku's mistakes myself. According to a check by Claude Opus 4.8, Sonnet got a quarter of the short dreams wrong. On dreams longer than about 150 words, it got 87 percent wrong. Each of these numbers was measured by a different checker from the same Claude family, and I don't have the sample sizes on record, so the tiers can't be compared directly.

In the end, not even Claude Opus 4.8 handled my rules. Its tags matched the plot, but not my conditions and the codebook's rules. So Claude Fable 5 redid all the tagging, and with its tags I was satisfied for the first time. A task that looks like sorting is, to me, really compression of meaning. No higher model checked the tags from Claude Fable 5.

A few things went wrong, too. For the final analysis, 45 agents running on Claude Sonnet wrote cards with quotes. On several cards, the agents didn't keep the dream numbers from the prompt and numbered them their own way. One of them quoted a sentence that couldn't be found in the archive. Claude Opus 4.8 then spot-checked forty more quotes and found nothing like it.

It was through quotes that weren't in the archive that a worse error came to light. The progress log claimed a batch of 30 dreams had been written in, when it hadn't. The write was most likely cut off when the credits ran out. Whether that one sentence nobody could find came from exactly this batch, I no longer know.

What came out of twenty years of dreams

Twenty years of dreams are readable and counted for the first time: over a hundred tags, statistics by year, what co-occurs with what. Claude Fable 5 wrote the final analysis on top of numbers from a script. But it was written before the new tagging, and whether its numbers hold on the new tags, I don't have on record.

Using the tags, the model could show how recurring themes changed over the years: some faded, others grew. In the end we also told dreams apart by their atmosphere, whether a dream was more positive or more negative, and that changed over time too.

But the main lesson, to me, isn't about dreams. “AI can't do it” and “AI can do it” are often statements about the prompt, not just about the model. Here both the model and the prompt changed, and I can't separate how much of the difference came from which. But the human in that prompt stood exactly where one is irreplaceable.

For me, the machine was the most meticulous scribe and librarian I have ever had. And the first reader who could say “I don't understand this.” In the first two cases where it said so, those were places where I was groping in the dark too. But it didn't replace my checking: elsewhere, the subagents made up dream numbering.

Model outputs are translated; the content is unchanged.