Google’s AI builds playable worlds in real time

Good morning, AI enthusiasts. The age of embodied AI is all about training in simulations…But what happens when some scenarios are too hard to construct, and even imagine?

Google’s recent Genie 3 just cracked that problem with the flexibility to generate wealthy, playable environments that evolve in real-time as AI agents (or users) explore them — unlocking a complete recent frontier of infinite training.

Share your input: We’re committed to repeatedly improving The Rundown, and we’d love your feedback to assist shape its future.

In today’s AI rundown:

• Google’s Genie 3 interactive world model

• OpenAI finally launches open-source models

• Turn any document into study video presentations

• Anthropic releases Claude Opus 4.1

• 4 recent AI tools & 4 job opportunities

LATEST DEVELOPMENTS

GOOGLE DEEPMIND

🌍 Google’s Genie 3 interactive world model

The Rundown: Google DeepMind just announced Genie 3, a brand new general-purpose world model that may generate interactive environments in real-time from a single text prompt, complete with surrounding and character consistency.

The main points:

• With Genie 3, users can generate unique, 720p environments with real-world physics and explore them in real-time, with recent visuals emerging at 24fps.

• The model’s visual memory goes up to 1 minute, enabling it to simulate the following scene while ensuring consistency with the previous ones.

• To attain this level of controllability, Google says, Genie computes relevant information from past trajectories multiple times per second.

• It also allows users to vary the worlds as they go by inserting recent characters, objects, or changing the environment dynamics entirely.

Why it matters: Genie 3’s consistent worlds, generated frame-by-frame in response to user motion, isn’t only a leap for gaming and entertainment. They lay the inspiration for scalable training of embodied AI, where machines can tackle the “what if” scenarios — like a path vanishing — by adapting in real time, identical to humans.

TOGETHER WITH SŌKOSUMI

🤖 The primary open marketplace for AI agents

The Rundown: Sōkosumi is the primary open marketplace for autonomous AI agents, letting you hire specialized AI co-workers for content, research, design, and data tasks with only a single click. Built for professionals and enterprise teams who need results — not subscriptions.

The platform offers:

• MCP ready, multi-model agents (ChatGPT, DeepSeek, Mistral)

• Enterprise-grade security with GDPR compliance and SSO integration

• Pay-per-task pricing — no monthly subscriptions or hidden fees

Construct your AI agent team today — redeem $100 in free credits with code RUN100.

OPENAI

🧠 OpenAI finally launches open-source models

The Rundown: OpenAI unveiled gpt-oss-120b and gpt-oss-20b, its long-awaited open-weight reasoning LLMs that match or exceed o4-mini and o3-mini in performance and can be found for local deployment under an Apache 2.0 license.

The main points:

• Available under Apache 2.0, the gpt-oss family, OpenAI’s first open LLMs since GPT-2 in 2019, immediately became #1 amongst 2M models on Hugging Face.

• The 120B variant performs on par with o4-mini on core benchmarks and exceeds on certain domains, while being deployable on an 80GB GPU.

• Meanwhile, the smaller 20B version is competitive with o3-mini, with suitability for local deployment on laptops with 16GB of memory.

• Each models feature adjustable reasoning (high, medium, low) and may handle agentic workflows, with function calling, web search, and Python execution.

Why it matters: After keeping its best models locked for years, OpenAI is finally living as much as its name, giving developers access to near-frontier reasoning models they will run and modify in their very own environments. It’s a serious boost for the open-source ecosystem, which has been rapidly closing the gap with closed models.

From the community: We’re hosting a workshop this Friday on the important thing advantages of local models and the way you possibly can run gpt-oss models in your computer.

AI TRAINING

🎥 Turn any document into study video presentations

The Rundown: On this tutorial, you’ll learn find out how to use NotebookLM’s recent ‘Video Overview’ feature to show documents into AI-narrated video presentations with slides — pulling images, diagrams, quotes, and numbers directly from the content.

Step-by-step:

1. Visit NotebookLM, create a brand new notebook, and upload your documents

2. Within the Studio panel on the suitable, click “Video Overview”

3. Optional: Click the three dots menu to customize your focus topics, audience, or learning goals

4. Review your generated video and “Download” to avoid wasting as MP4

Pro tip: Create multiple Video Overviews in a single notebook, make versions for various audiences, or give attention to different chapters of your content.

PRESENTED BY LOVART

🎨 Your next creative partner won’t be human

The Rundown: Lovart has officially exited beta with its AI design platform built for visual collaboration. It incorporates a creative reasoning agent who thinks with you, sources references, and builds brand systems in minutes. Designers say it feels more like working with a teammate than using a tool.

With Lovart, you possibly can:

• Use the brand new ChatCanvas feature to visually collaborate along with your agent, iterate, and refine in real time through natural language.

• Turn easy prompts into brand visuals, social content, videos, and even 3D models

• Depend on a specialized agent to take care of style consistency across materials

• Work faster with memory that learns the way you design and adapts to your habits

Transform your creative process with the world’s first design agent today.

ANTHROPIC

🤖 Anthropic releases Claude Opus 4.1

The Rundown: Anthropic released Claude Opus 4.1, an incremental upgrade for Opus 4, improving performance across real-world coding, in-depth research, and data evaluation tasks — particularly those requiring attention to detail and agentic actions.

The main points:

• Opus 4.1 brings a notable coding upgrade over its predecessor, boosting performance on SWE-bench Verified from 72.5% to 74.5%.

• Improvements are also seen across math, agentic terminal coding (TerminalBench), GPQA reasoning, and visual reasoning (MMMU) benchmarks.

• Customers cited real-world gains with the model, saying it excels at tasks like multi-file code refactoring and identifying correlations in codebases.

• Anthropic said the upgrade—available to paid users and businesses—marks the beginning of “substantially larger improvements” planned for its models.

Why it matters: With Opus 4.1, Anthropic is adding more momentum to what’s shaping as much as be an exciting week for AI enthusiasts. The upgrades are a welcome gift, but with OpenAI’s GPT-5 potentially dropping any day, all eyes will probably be on how the corporate’s models hold its ground, especially in coding, where it has stood out.

QUICK HITS

🛠️ Trending AI Tools

• 🔄 Depot’s Claude Code Sessions – Persistent AI coding sessions that sync across teams & environments for seamless collaboration*

• ⚙️ Kaggle Game Arena – Benchmark to check LLMs on evolving strategic games

• 📽️ ChatGPT – OpenAI’s AI assistant, now with tools to detect mental distress

• 📝 Gemini Storybooks – Google’s AI now creates narrated storybooks

^{*Sponsored listing}

💼 AI Job Opportunities

• ⚙️ The Rundown – Growth and Content Strategist

• 📢 Groq – Product Marketing Manager, Sales Enablement

• 🛠️ Figure AI – Prototype Development Technician

• 🎭 Meta – Creative Director

📰 All the things else in AI today

ElevenLabs introduced Eleven Music, its multilingual music generation model with control over genre, style, and structure, and the choice to edit each sounds and lyrics.

Google added a brand new Storybook feature to the Gemini app, allowing users to generate personalized storybooks about anything with read-aloud narration without cost.

Perplexity acquired Invisible, an organization developing a multi-agent orchestration platform, to scale its Comet browser for consumer and enterprise users.

Elon Musk shared Grok’s Imagine image and video generator is seeing massive interest, with 20 million images generated yesterday alone.

Alibaba released its Flash series of Qwen3-Coder and Qwen3-2507 models via API, with as much as 1M-token context window and low pricing.

Shopify added recent agent-focused features, including a checkout kit to embed commerce widgets into agents, low-latency global product search, and a universal cart.

COMMUNITY

🎥 Join our next live workshop

Join our next workshop this Friday, August eighth, at 4 PM EST with Dr. Alvaro Cintas, The Rundown’s AI professor. By the top of the workshop, you’ll know hot to run your personal fully private, open-weight model locally in your computer.

RSVP here. Not a member? Join The Rundown University on a 14-day free trial.

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer—the humans behind The Rundown