SIMA 2: An agent that plays, reasons, and learns with you in virtual 3D worlds

Original Article ↗ Hacker News Discussion ↗

Architecture, Gemini & Demo Authenticity

  • Commenters infer SIMA 2 is a separate agent layered on top of Gemini, interacting via a text interface.
  • Some scrutinize the demo video, pointing to a slight grammatical mismatch in the on-screen “reasoning” text as evidence the captions may be post-produced rather than raw model output. Others argue the context (“ripe tomato” text seen earlier) explains the phrasing and think the marketing is reasonable.

Game Worlds vs World Generation

  • Several people are confused by the video and blog as to what is generated. Clarification in the thread: SIMA 2 is a game-playing agent; most of the demo is just No Man’s Sky, not a SIMA-generated world.
  • Genie 3 is mentioned separately as Google’s world-model / world-generating line of work.

Performance, Generalization & ‘True Intelligence’

  • Some are impressed by reported 65% success on all tasks and especially ~15% on unseen environments, seeing it as a big leap over recent “LLM plays games” efforts.
  • Others emphasize how low 15% is and call the charts misleading, arguing this is still far from being broadly useful.
  • There is debate about “true intelligence”: some see large-scale task coverage as the only realistic path, others stress humans’ superior zero-shot reasoning and point to domains where AIs still lag.

Robotics, Sim2Real & Control Abstractions

  • Several comments connect SIMA 2 to robotics: high-level agents issuing low-dimensional commands (“move here”, “empty the dishwasher”) to lower-level control systems that handle physics and actuation.
  • Skeptics note that real-world robotics is hard due to occlusions, unactuated objects, adversarial agents, and safety constraints; progress may require more than just more data.
  • The sim-to-real transfer problem is highlighted; SIMA-style work is seen as groundwork to be combined later with higher-fidelity world models and physical robots.

Openness, Research Lineage & Dreamer

  • Some wish Google would return to more open-sourcing, contrasting current polished blog posts with earlier releases.
  • Dreamer v3/v4 and Minecraft agents are referenced as related open research in model-based RL and offline training.

Use Cases: Agents as Helpers, NPCs & ‘Gaming Minions’

  • Many imagine agents as cooperative partners: handling grind, acting as co-op companions, or populating game worlds with more intelligent NPCs.
  • Others find the idea of AI playing games for you anticlimactic or tantamount to cheating, especially in grind-based MMOs.
  • There is enthusiasm for SIMA-like systems as fast “computer use” agents (mouse/keyboard at high FPS, phone automation), which current tools lack.

Impacts on Games, E-sports & Society

  • Some worry about AI ruining online games and e-sports via unbeatable bots and 24/7 farming; others compare this to chess, where human competition persists despite stronger engines.
  • A few comments zoom out to broader concerns: AI making many humans economically “irrelevant,” skepticism about narratives like universal basic income, and fear that advanced agents primarily enrich those who already control capital.