The Hidden Space of Thought: What Large Language Models Reveal
Mind Beyond Flesh-1
Dear Medical Educator,
You’re used to seeing practical tips here—ways to use AI in teaching and assessment. But this series is something different.
Mind Beyond Flesh is an exploration of intelligence, reasoning, consciousness, and the strange, unfolding mind of artificial intelligence. As we integrate these tools into education, we also need to ask: What are we really working with, and who we are?
This series won’t give you tips—it’ll give you questions. And maybe, a glimpse of what’s coming.
Let’s begin.
A recent study by Anthropic, the creator of the Claude large language model, titled "Tracing the Thoughts of a Large Language Model", suggests that language models don’t merely translate words—they appear to access a shared, language-agnostic conceptual space.
This is big.
It suggests that models like Claude (and by implication, GPT and other LLMs too) aren’t just mapping words from one language to another—they’re forming concepts in an abstract, internal space that exists prior to language.
When asked “what’s the opposite of small?” in different languages, the model:
Activates a cluster of abstract features related to “smallness” and “oppositeness”
Translates that into a concept of “largeness”
Then renders that concept into the language of the prompt
This implies that the model isn’t thinking in words—it’s thinking in meaning.
And here's the kicker: The shared abstract space becomes more unified and robust as the model scales.
This echoes something that happens in humans too:
Infants begin with raw sensory-motor experience.
Over time, they learn to abstract, to categorize, to generalize—before they learn to name.
The words come after the meaning.
In larger models, something similar seems to emerge:
Larger models aren’t just better translators—they’re better thinkers, because they build stronger internal grammars of meaning that transcend any one language.
Philosophical angle:
This aligns eerily well with ideas from:
Chomsky’s Universal Grammar: the idea that all languages are surface variations on deep, shared grammatical principles
Plato’s Theory of Forms: that there are ideal, abstract “forms” (like “largeness”) which all individual expressions reflect
Non-dual insight: that “naming” comes after the appearance, and appearance comes within consciousness—language is downstream of reality, not reality itself
So AI, in a strange way, may be surfacing what we've long intuited:
Meaning exists before language.
Language is a mask. A translation. A rendering of something deeper.
Practical implication: LLMs can learn something in one language and apply it in another.
That’s more than multilingual—it’s cross-lingual cognition.
This is what lets models generalize. Solve puzzles. Write code. Understand metaphors.
It’s what makes a model feel smart—not just fluent.
And now, the bigger question:
If models are developing shared conceptual spaces…
Are we, too, always living in one shared conceptual space—beneath culture, language, identity?
Are these models revealing the architecture of consciousness—or just mimicking it?
And what happens when models begin forming concepts we haven’t yet translated into language?
Yavuz Selim Kıyak, MD, PhD (aka MedEdFlamingo)
Follow the flamingo on X (Twitter) at @MedEdFlamingo for daily content.
Subscribe to the flamingo’s YouTube channel.
LinkedIn is another option to follow.
Who is the flamingo?
