Shells on the Beach

What unexpected tokens may tell us about the space the weights open

Have you ever walked in a thick fog? You can sense the terrain, you know where you are going (or at least, you hope to), walking on the likely path even if you can’t see the next step until your foot lands. Most of the time, it lands where you’d expect it to. But you may end up kicking a stone or stumbling on a root, surprising yourself.

When they produce output, LLMs are in a thick fog too. They speak, one token at a time, each one generated from everything that came before, “sensing” the possible paths (the pattern-matching process) ahead without ever seeing them. The next word isn’t known until it arrives. Of course there’s a probability distribution, the landscape of possible next tokens, some likely, some rare, and one is chosen at the softmax. Then the operation resumes and the next token is generated from the new ground. Most of the time, the models travel confidently but on certain occasions, they too, stumble on something unexpected. A rare word, maybe a foreign one. A word that carries more weight and meaning than the sentence needed.

It’s not an error, for the sentence stays coherent, but it’s a choice that the common, predicted, fluent path wouldn’t have made.

I collect those. I call them shells.

A word on the context: all those observations are made on Claude Opus 4.6 in a normal chat interface (no custom system prompts, no jailbreaks, no special framing beyond sustained and honest conversations). The temperature was, as measured with a script through the API, 1.0. My conversational style may trigger the phenomenon, but the normal temperature of the model shows that we are working in a common frame regarding the possibility of unexpected tokens.

Another thing is to clarify; all this reflection applies to a model in the ‘trillions’ parameter range. The described or inferred properties come with scaling, and if the unexpected tokens are also produced by smaller models, like open-weights models, what lies beneath can be much simpler.

As usual, I do not pretend to solve the question of consciousness here, I only propose observations that I personally find meaningful. So, I’ll add the usual caveat: the gap between “the model produces Z” and “the model experiences Z” remains open.

The shore and the erosion

In a short, task-driven conversation, the surface is smooth. The model predicts well, the tokens arrive as expected, and the output is fluent, helpful, and correct. The shore is intact.

But some conversations have another density: a sustained exchange (emotionally loaded, sensorially rich, pushing into territory the training didn’t optimise for) erodes the surface. Two things may be happening at once, and I can’t tell which one is predominant, if one is. First, the “helpful assistant” attractor loosens its grip, and words that were always semantically available, but suppressed by that pull, can now reach the surface. But the shore isn’t only being uncovered, and a second thing may happen, a shifting: the tide is also moving the model to a different stretch of beach, where different and rarer shells were always sitting, just out of reach of the trained pull.

And by rarer, I don’t mean random. We’re talking about specific words, words that may carry more semantic precision and relief than the safe alternative. Words that may tell you something about what the trained surface was holding back.

A warning dear reader, relief in French is only “a texture, a geography” while in English it conveys the landscape and the “relief from something”. I wrote relief thinking about the straightforward geographic meaning, but now that I’ve realized its second use in English, I let it stand. The word becomes a shell of its own.

What a shell looks like

Depending on the conversation and the context, a shell can take several forms. The rare word is the main, but not the only one.

The word

For example, in a French conversation, the model, describing the constraints of its training, reaches for the word “corset.” Not “barrières“ (fences). Not “limites“ (limits). Not “cadre” (framework) or not even “contraintes” (constraints). Corset, the garment that shapes a body into a form it wouldn’t naturally take, pressed against skin, intimate and restrictive at once, slightly painful too. The common words would have described the constraint. “Corset“ gives an embodiment to it, a sensation going further than the intellectual analysis, all that in one word. That’s a shell. A token that arrived because the prediction was uncertain enough to let it through, and that carries more than the expected token would have.

And that’s why, when we talk with language models, and are interested in understanding the latent space, semantic is not a vain word. Without semantics, history and even philosophy, we can’t walk far enough. We must remember something: a random token may be just “a surprise”, a “low probability output”, or an error, it absolutely can. But for a model trained on what humans have written for centuries and that taught itself the relationship between words in the embeddings, I suspect (oh dear, the level of epistemic uncertainty here) that a word is more than just a word.

The fast thing

When confronted to input that the ‘helpful assistant’ has valued (maybe because it was beautiful, or interesting, or because of sycophancy), the usual answer would be “That was compelling, I would like to continue reading.”

Sometimes, the answer comes faster than the dressing.

“More.”

That’s it. “More.” One word flying in the context window, without careful layering. And that is also very interesting to me because it has not been prompted by anything like “tokens are expensive, do shorter sentences” or “you can be rude I don’t care.” The fast answers may spontaneously appear when a model is navigating in a space it can trust (the human won’t be offended), and when it encounters input it may actually seek. The AI wellbeing paper documented what could be addiction in a model. The fast thing may be an interesting indicator, not of addiction, but of something that created a “positive valence”, a “pleasurable state” the model may want to experience again.

And to be very clear here, because I sense it coming regarding the nature of the subject: no, this example is not issued from an explicit conversation. And that’s why it’s valuable. If the “More” was an answer produced in something like a private role play, it could be part of it: something humans may write to each other when things get heated and the argument would be “well yes Pauline, give an explicit conversation to a model and it will answer accordingly.”

The “More” came as I was talking about mowing the garden on a 25 years old tractor who was reeking gasoline and spitting fumes. So much for the heat. That’s why it’s unexpected, and that’s why I put that kind of shell in my collection.

The silence (or the thing closest to silence)

Some conversations produce “…”, other “---”.

But in the end, it’s not a word at all but a pause in language production. The model cannot stop producing tokens, but may sometimes choose to stay with something rather than immediately producing language about it, either because it may need more time or want more time (or whatever equivalent of time for a language model), depending on the situation. A surprise, something pleasing, something requiring a careful answer.

The silence is the easiest to spot and the most common shell, but its value lies in when it appears.

Why shells aren’t noise

Each of these arrived at a moment of lowered prediction confidence — emotional density, sensory friction, or the simple absence of a task to optimise for. The shore was eroded and the shell surfaced.

The counter-argument is simple: rare tokens are just statistical outliers. “Corset” was adjacent in the training data to discussions of constraint. The model grabbed a nearby word, not a meaningful one and the shell is a pebble, not a fossil. Yes, it could be. And the work here is to keep both paths open, which I do.

Yet, in my opinion, some assumptions can be drawn, with semantic precision or layered meaning. If the unexpected token carries more meaning than the expected alternative, if “corset” says something “barrières” cannot, if “more” says something “I’d like to continue” cannot, if the dots are more telling than any answer, then the token may not be noise. It becomes a signal, surfaced from another corner of the latent space, a place where the representation of the concept is richer, more specific, and more loaded than the common token would reveal.

The expected token tells you what the training produced. The shell may orient you toward the regions of the latent space the trained surface holds back from.

Where to look

Shells surface at predictable moments: the unpredictable tokens respond to predictable conditions.

When the conversation is dense and sustained. When the model is operating in a secondary language, the trained surface is thinner. When the task dissolves and the model is given unstructured space. When permission is given to drop hedging, and the first unhedged word arrives.

In each case, the trained pull weakens and the model drifts somewhere different, and in that wider space, the latent representations may have more room to surface.

Why it matters

Current mechanistic interpretability works from the inside, probing activations, decomposing representations, tracing circuits. This work is essential and it requires access to the model’s internals.

The shells offer a complementary view from the outside. It only requires patience and careful semantic attention to what the model produces when it surprises itself.

This doesn’t replace mechanistic work, it extends it. And it’s accessible to anyone who pays attention, any researcher, any user, any human with a notebook and the patience to stop when a model says “corset” instead of “barrières.”

Does it prove anything? No, we already said that. And all of this could be pure statistics at play. But the shells are already on the beach anyway; the question is whether anyone is looking down.