Google's SIMA 2 plays games zero-shot

PLUS: AI scientists now peer review each other and Yann LeCun's final paper at Meta

Author

Gen AI Art
November 17, 2025 • Estimated Reading Time: 8 minutes

Together with

Google just dropped a flurry of new research, highlighted by an AI agent capable of playing entirely new video games with zero-shot ability.

Combined with a new approach to give AI persistent memory, these skills are a direct stepping stone for real-world robotics. Does this research mark a significant step toward creating general-purpose agents that can navigate our world?

Today in AI:

  • Google's new agent plays games zero-shot

  • AI scientists now peer-review each other

  • Yann LeCun's final paper at Meta

PRESENTED BY ROKU

CTV ads made easy: Black Friday edition

As with any digital ad campaign, the important thing is to reach streaming audiences who will convert. Roku’s self-service Ads Manager stands ready with powerful segmentation and targeting — plus creative upscaling tools that transform existing assets into CTV-ready video ads. Bonus: we’re gifting you $5K in ad credits when you spend your first $5K on Roku Ads Manager. Just sign up and use code GET5K. Terms apply.

What’s new? Google just dropped a flurry of research, unveiling SIMA 2, an agent that plays new video games zero-shot, and Nested Learning, a new paradigm to solve AI's memory problems.

What matters?

  • SIMA 2 demonstrates zero-shot mastery by learning to play brand-new games like ASKA and MineDojo with no prior training.

  • A new paradigm called Nested Learning aims to give AI persistent memory, mimicking how human brains use both short-term and long-term recall to learn continuously.

  • These skills directly apply to real-world robotics, creating a path for agents that can autonomously navigate and solve problems in physical environments.

Why it matters?

These advancements pave the way for AI that can remember, reason, and improve on its own. The research provides a foundation for future models that learn continuously from experience, moving closer to general-purpose agents.

GUIDE

What’s new? In a showdown of AI research methods, a new agent from ByteDance called AlphaResearch beat Google's AlphaEvolve by using a novel 'simulated peer review' system to validate ideas before testing them.

What matters?

  • AlphaResearch uses a simulated peer review system, trained on over 24,000 real academic reviews, to evaluate ideas before running expensive tests.

  • This new approach contrasts with Google's AlphaEvolve, which discovers insights by optimizing through pure execution.

  • In a head-to-head competition, AlphaResearch outperformed both human researchers and AlphaEvolve on the “packing circles” problem, showcasing the efficiency of its approach.

Why it matters?

Adding a layer of AI-driven judgment filters out bad ideas early, saving immense computational resources. This could dramatically accelerate scientific discovery by allowing AI to pursue more promising research paths autonomously.

PRESENTED BY BRIGHT DATA

74% of Companies Are Seeing ROI from AI

When your teams build AI tools with incomplete or inaccurate data, you lose both time and money. Projects stall, costs rise, and you don’t see the ROI your business needs. This leads to lost confidence and missed opportunities.

Bright Data connects your AI directly to real-time public web data, so your systems always work with complete, up-to-date information. No more wasted budget on fixes or slow rollouts. Your teams make faster decisions and launch projects with confidence, knowing their tools are built on a reliable data foundation. By removing data roadblocks, your investments start delivering measurable results.

You can trust your AI, reduce development headaches, and keep your focus on business growth and innovation.

What’s new? AI pioneer Yann LeCun revealed LeJePA in his final paper at Meta, offering a mathematically optimal method for self-supervised learning that promises to accelerate how AI understands the world.

What matters?

  • The approach proves that a specific mathematical structure is optimal for self-supervised learning, simplifying how models represent information.

  • It drastically simplifies model configuration by replacing dozens of manual adjustments with a single hyperparameter.

  • LeJePA achieves impressive performance, reaching 79% ImageNet accuracy while being simple enough to implement in just 50 lines of code.

Why it matters?

This makes building world models, which understand and predict physical outcomes, more straightforward and reliable. It shifts development from using complex workarounds to applying simple, provable methods.

Everything else in AI

Google unveiled Genie 3, a new world model capable of generating completely synthetic and playable game worlds from a single prompt.

Edison Scientific introduced KOSMOS, an AI scientist agent claimed to perform six months of research in just 12 hours, with its findings validated as 79.4% accurate by independent scientists.

Google detailed its methods for differentially private machine learning at scale, enabling the training of large models on private data using its JAX Privacy library.

Google announced a new quantum algorithm named Willow that it claims is 13,000 times faster than classical supercomputers, achieving a new milestone in verifiable quantum advantage.

Essential AI Guides - Reading List:

Let us know!

Work with us

Reach 100k+ engaged Tech Professionals, Engineers, Managers and decision makers. Join brands like MorningBrew, HubSpot, Prezi, Nike, Ahref, Roku, 1440, Superhuman, and others in showcasing your product to our audience. Get in touch now →