Recorded and played back. Switch tabs to see GPT-2 with nothing behind it.
On its own, GPT-2 wanders and loses the thread within a sentence. Apex holds one line of thought across the whole exchange. The telling part is the vocabulary: almost none of the words it uses appear in the novel it trained on. That rules out the easy explanations. It is not recalling text, retrieving passages, or matching on semantic similarity. It learned the structure beneath the words, and that structure, run through GPT-2's vocabulary, comes out as language it was never shown, from roughly 150,000 tokens, nearly five orders of magnitude less data than a model like GPT-2 is trained on.
It does this with no context window and no attention. The topic is not held in a buffer and re-read each step; it is a stable state inside Apex, carried forward on its own. And that state was sharp enough to steer a frozen model Apex had never trained with, which means the representations are general and transportable. Align the decoder, and what comes out is fluent language, backed by the memory and reasoning already inside Apex. Apex is the mind. Any model can be the voice.
Language is just one modality to Apex, and it emerges on its own. Bringing it to fluency is ongoing research, and we are building the physical infrastructure for it. What you see here is a research-level result, not a product.
Early internal research. Newer results are private. How Apex works.