January 28, 2026
Autocomplete grows up
World Models
Big AI names bet on ‘world models’—and the comments are chaos
TLDR: Big labs are racing to build “world models” that predict what happens next in real situations, not just the next word. Comments explode into a semantics fight (transformer vs GPT), skepticism about whether models already simulate events, and jokes about rebranding recommendation engines—signaling a big, controversial shift.
The AI world is buzzing: Yann LeCun reportedly striking out to build a lab around “world models,” Google dropping Genie 3, Demis Hassabis hinting he’s spending most of his research time here, Anthropic showing models already form fuzzy internal “worlds,” and OpenAI calling Sora a world simulator. Meta even has a code-focused version where a smaller model beat bigger ones on tests that require running code, not just looking at it. Translation for non-nerds: instead of guessing the next word, these systems try to predict what actually happens next in a situation.
But the comments? Pure spectacle. One camp cheers, saying this is finally simulation over autocomplete, while skeptics clap back: “Don’t confuse transformers with GPTs!” and argue current models already learned how events unfold. A spicy thread accuses “world models” of being rebranded recommendation engines—systems that predict how you’ll react when shown something—aka simulating people for clicks. Others crack jokes: “Genie 3: make a wish, get a world,” and “World Model Cinematic Universe, Phase One.” The oddball comment about “Cloudflare, Palo Alto Networks… alternative charset” becomes a meme for broken simulations (“NPC error: dialogue corrupted”). Whether it’s science revolution or relabeling party, the vibe is hype vs side-eye, and everyone’s rolling the dice on the next big paradigm.
Key Points
- •The article asserts major AI labs are simultaneously focusing on world models, citing multiple examples across organizations.
- •A world model is defined as predicting the next state or observation, capturing causal laws of an environment, distinct from next-token and reward-optimized models.
- •Meta’s September paper on a 32B Code World Model reports performance matching or exceeding larger models on benchmarks like SWE-Bench and Terminal Bench.
- •OpenAI positioned Sora as a world simulator, Google announced Genie 3 for world simulation, and Veo3 is described as a physics/world model.
- •The article claims world-model-like systems already exist in practice (e.g., recommendation engines, trading, supply chains, weather, game engines) because they predict states rather than tokens.