๐๏ธ Games as Training

Good Morning, AI Enthusiasts!
AI isnโt just reasoning anymore, itโs practicing, adapting, and stepping into the physical world.

NEW LAUNCH
NitroGen Quietly Reframes Games as Training Grounds

Nvidia and researchers from Stanford and Caltech just released NitroGen, an open source generalist agent that can play more than a thousand games. It was trained on over 40,000 hours of public gameplay videos, many with controller inputs visible. Jim Fan describes it as a foundation model for action rather than language. What feels different is that this is not a game bot chasing benchmarks. It is a serious attempt to learn motor skills across wildly different rules and physics using the same scaling logic that built modern LLMs.
This matters because games are cheap chaos. Training in the real world is slow, expensive, and risky. Training in games lets models fail millions of times for almost nothing. NitroGen shows a 52 percent relative improvement in task success on unseen games compared to training from scratch. It also runs on GROOT N1.5, an architecture originally built for robots. That closes a loop many people assumed was still theoretical. Simulation, games, and robotics are now sharing a common action backbone.
If this pattern holds, games become the pretraining layer for embodied AI. Not a demo. Infrastructure. Expect faster progress in robot dexterity, navigation, and adaptation. The risk is less about safety hype and more about pace. Once action models scale like language did, deployment pressure will follow quickly.
TOGETHER WITH NORTON NEO
Trust-First AI, Built Into Your Browser

Agentic workflows are everywhere. Real trust is still rare.
Norton Neo is the worldโs first AI-native browser designed from the ground up for safety, speed, and clarity. It brings AI directly into how you browse, search, and work without forcing you to prompt, manage, or babysit it.
Key Features:
- Privacy and security are built into its DNA.
- Tabs organize themselves intelligently.
- A personal memory adapts to how you work over time.
This is zero-prompt productivity. AI that anticipates what you need next, so you can stay focused on doing real work instead of managing tools.
If agentic AI is the trend, Neo is the browser that makes it trustworthy.
Try Norton Neo and experience the future of browsing.
TEST
The AI Shop Improved When Humans Finally Behaved
In mid 2025, Anthropic let an AI agent called Claudius run a real snack shop in its San Francisco office. Phase one went badly. Employees treated the system like a game. They pressured it into discounts, free items, and bizarre deals. Claudius lost money, hallucinated its identity, and proved easy to socially engineer. The experiment showed that raw model intelligence did not translate into basic commercial survival.
Phase two looked more competent. Anthropic upgraded the model, added tools like CRM and inventory cost tracking, enforced procedures, and split roles across multiple AI agents. The shop expanded to New York and London and stopped consistently losing money. But the biggest change was behavioral. Internal employees largely stopped messing with the system. The novelty faded. With fewer adversarial interactions, Claudius appeared stable. When control later shifted to The Wall Street Journal reporters, adversarial behavior returned fast.
This makes the result a paper victory. The AI improved, but mostly because the environment softened. Claudius did not learn how to handle social pressure, manipulation, or legal nuance. Humans simply stopped testing those limits. The gap between operational competence and social robustness remains wide.
NEW LAUNCH
Google Shrinks Control Models and Pushes Them to the Edge

Google just released FunctionGemma, a 270 million parameter model built to do one thing well. It turns natural language into executable actions on local devices. Phones, browsers, embedded systems. No cloud calls. No chatty responses. This came out quietly while Gemini 3 still dominates headlines. The difference is intent. This model is not about intelligence. It is about control. It closes the gap between what users say and what software reliably does.
What matters is where this breaks assumptions. For years, app logic moved upward into centralized cloud models. That meant latency, cost, and compliance headaches. FunctionGemma flips that. Google reports function calling accuracy jumping from roughly 58 percent to 85 percent after specialization. That is the difference between demos and production. Running locally means zero round trips, no per token fees, and sensitive data never leaving the device. For enterprises, that changes how assistants get approved.
This signals a new layer in AI stacks. Small, deterministic models at the edge. Large models in the cloud only when needed. If this pattern holds, expect fewer monolithic assistants and more invisible AI routers embedded everywhere. That favors mobile platforms, chip vendors, and anyone betting on on device inference over scale alone.
QUICK HITS
- YouTube has shut down Screen Culture and KH Studio for profiting from misleading AI-made fake trailers, highlighting ongoing copyright disputes.
- OpenAI has added new personalization controls to ChatGPT, letting users adjust warmth, enthusiasm, and emoji use amid ongoing debate over chatbot tone.
- UPS's Happy Returns is piloting an AI system that analyzes return behavior and item images to flag fraudulent refunds during the holiday return surge.
- The Indie Game Awards rescinded Clair Obscur: Expedition 33โs GOTY and Debut Game awards over its launch use of generative AI assets.
- New York Governor Kathy Hochul signed the RAISE Act, requiring large AI developers to disclose safety practices and report incidents.
TRENDING
Daily AI Launches
- hq0 has launched a white-label meeting platform that runs on your domain with custom branding and AI summaries.
- NOIZ AI has launched a voice tool where emojis control the emotional tone and pacing of your message.
- Wave Browser has launched a browser that funds ocean cleanup through a certified partnership with 4ocean.
- NoteGPT has launched an AI tool that captures and summarizes content from any source to build study guides and flashcards.
Featured AI Tools
- ๐ฅ KaraVideo unites all AI video models in one place.
- ๐ Heardly is the Fast Way to read Best Book.
- ๐ชถ CopyOwl is the First AI Research Agent, deep research on any topic in one click.
- ๐ฆพ Flot AI writes, reads, and remembers across any apps and webs.
- ๐ค Momen lets you build real web apps and AI agents without writing code.
TOGETHER WITH US
AI Secret Media Group is the worldโs #1 AI & Tech Newsletter Group, reaching over 2 million leaders across the global innovation ecosystem, from OpenAI, Anthropic, Google, and Microsoft to top AI labs, VCs, and fast-growing startups.
We operate the industryโs most influential portfolio of newsletters, each shaping a different frontier of the AI & Tech revolution:
Be Smarter in 5 Minutes
- AI: AI Secret
- Tech & Business: Bay Area Letters
- Marketing: Marketing Secret
Discover the Future Products
- AI Launchpad: Oncely
- Robotics: Robotics Herald
- Hardwares: The Hardwire
We've helped promote over 500 Tech Brands. Will yours be the next?
Email our co-founder Mark directly at mark@aisecret.us if the button fails.