Because I often, even in this sub, read the notion that "LLMs are not enough for world models and embodiment" and "JEPA this, GENIE that, blah blah blah," the good folks at Tencent have come to my rescue and validated what I've been preaching ever since I uploaded my first pron model to Civitai:
LLMs are enough.
Tencent basically built "Anti-JEPA" and proved that LeCun is, once again, full of sht.
LeCun wants a silent latent physics engine with an LLM bolted onto the side. Tencent has built an LLM/VLM that can reason in language and literally dream its future visual states inside the same autoregressive stream. Whether it works reliably is still an open question, since it is the first of its kind, but conceptually it looks like yet another demonstration that you can keep extending LLMs until they eat the architecture that was supposed to replace them.