Large language models (LLMs) like chatGPT or DeepSeek AI essentially function via probabilistic next-element prediction. That is, they predict the next word in a sequence based on statistical patterns learned from huge datasets. After waking up from a pretty bizarre dream this morning,  I thought about odd similarities between LLM confabulations and dreams, both of which appear to produce locally coherent but globally bizarre output.

LLM feverdreams in 2023

When these models came out, they regularly confabulated. In January 2023, I wrote a short blog post on asking chatGPT to provide some scientific references for me. It did—plausible references including plausible journal and author names—but these references did not exist. They could have, though—some of them were pretty close to the real targets. But they didn’t.

Interestingly, when I confronted chatGPT with the potential that these references weren’t real, it doubled down and told me that they very much were real.

LLMs have much improved in the 18 months since then, but they still confabulate. And those examples of confabulation have helped me a little to understand how LLMs generate words. This week, for example, I asked it to create statistical syntax for a figure, which it did, choosing 3 colors. I then asked it why it chose these colors, to which it gave a fairly surprising answer: because the colors matched the psychological constructs in the plot.

When confronted, unlike 18 months ago, it freely admitted that this just wasn’t true. Some progress, I guess.

LLMs and dreams: prioritizing local over global coherence

LLMs try to predict the next word in a sentence based on learned patterns, and this can be done in many different ways. Some output preserves local coherence: it is a word that commonly follows another word. Some output preserves global coherence: it makes sense in this sentence or story. When LLMs confabulate, it appears they maintain local plausibility over global consistency.

• Both LLMs and dreams appear to perform sequence prediction based to some degree on prior experience or information.
• They are both trained on data, which can lead to some form of novel recombination. In a way, we could say that both systems hallucinate plausible sequences from past information.
• Both systems confabulate, and maybe the reason is that they both produce locally coherent outputs that at times are globally odd. It is rarely the case that the directly next step in any dream by itself is bizarre: it is, I would argue, usually locally coherent. Things as a whole can be bizarre, seen from a distance.
• I’m neither neuroimaging nor LLM expert, but my understanding is that human dreaming is at least thought to involve spreading activation across neural networks during sleep, especially in REM. This isn’t entirely different from how LLMs activate words or tokens based on contextual embeddings and associative similarity.

Short random blog post in the hopes to start discussions on this. Would love to learn more about the neuroscience of dreams, in case there is robust work on the topic.