- NATURAL 20
- Posts
- AI Moves From Answers to Action
AI Moves From Answers to Action
PLUS: Apple clears a China hurdle, Europe opens Android to AI rivals, and Nvidia expands its Japan robotics push.
Wes Roth
July 17, 2026
Sponsored by
Your Website Shouldn't Look Like It Was AI-Generated
You know the pain. Your AI builder ships something functional. Technically a website. But your customers can tell.
Readdy.ai generates designs that actually convert, with clean layouts, professional copy, real visual hierarchy, and distinct custom designs that don't look like AI.
Get a site that's as special as your business. Live in under 5 minutes.
Agency quality. DIY price.
Today:
AI Mode Connects Apps to Complete Everyday Tasks
Build Turns Phone Prompts Into Playable Games
Gemini Omni Flash Brings Prompt Editing to Vids
Apple Intelligence Clears a Major China Hurdle
Android Opens Core Features to Rival AI Assistants
AI Moves From Answers to Action
Google AI Mode acts across connected apps, Roblox builds games from phone prompts, and Gemini Omni Flash edits finished video inside Google Vids.
AI is becoming part of the product interface, not a separate destination.
The newest releases let people move from a request to a grocery cart, a playable game, or an edited video without stitching together several tools.
Google is rolling out connected apps inside AI Mode in Search, starting this week in the United States. Users can securely link Instacart, Canva, and YouTube Music, then ask Search to act across those services.
A grocery request can add ingredients to an Instacart cart, although checkout still happens in the Instacart app or website. Search can also surface Canva templates and create, save, and play a YouTube Music playlist.
The integrations can work with Personal Intelligence, Google’s system for tailoring answers from a user’s connected information. Google says more partners are coming, but it did not announce a wider rollout date or separate price.
The benefit is fewer handoffs between search and another app. The tradeoff is that useful personalization requires people to connect more services and decide how much account information they want Search to use.
Roblox announced Build, a mobile-first creation tab that turns a text prompt into a basic playable game inside the Roblox app. It can generate gameplay rules, environments, characters, visual style, sound, and working 3D scenes.
Creators can refine the result, test it, share it, publish it, or continue the same project in Roblox Studio. Roblox says Build combines its own models with open-source systems and uses game-specific training data.
A public test begins July 28 in New Zealand for age-checked users nine and older. Published games that pass Roblox’s safety checks will be available globally to age-checked users 16 and older.
The base version will be free, with paid options planned for advanced users. Roblox is also developing agents for playtesting, analytics, and experiments, but those tools do not yet have firm release dates.
Roblox says Build-made games will face the same extended review and engagement-based discovery system as other games. That may limit low-quality output, but the effect on moderation and the volume of generated games will only become clear after testing.
Google launched Gemini Omni Flash in Google Vids, giving paid Workspace users a way to edit video with plain-language instructions. Google says Vids now has seven million monthly users.
Users can change lighting, remove clutter, replace backgrounds, adjust product colors, alter camera angles, and update voiceovers or on-screen text. The same model can create new clips with a personal avatar that looks and sounds like the user.
Personal avatars require a secure face-and-voice capture, are limited to users 18 and older, and currently use English as the primary language. Workspace administrators can control access, and Google did not announce a separate price beyond an eligible paid Workspace account.
Google adds an invisible SynthID marker to generated videos so its systems can identify AI-made content. The company did not publish independent quality tests, and users still need to review edits, identity likeness, and factual claims before sharing a finished video.
🧠RESEARCH
Researchers propose Deep Interaction, which lets people edit the faulty part of an AI model’s reasoning while preserving correct steps. On science and mathematics tasks, the method improved correction success by more than 25% and cut text processing by about 40% versus ordinary follow-up prompts, according to the authors’ tests.
Researchers tested whether improvements to AI agents survive new tasks. All three methods improved, but only RELAI-VCL kept learning without losing gains. It achieved a 76.4% average pass rate, ahead of GEPA at 66.0%, Meta Harness at 64.6%, and the unchanged baseline at 58.7% across evaluations.
Five experiments involving 3,132 people found that access to wrong AI advice erased participants’ willingness to answer “I don’t know.” They answered more questions but were correct about one-third as often, while confidence nearly doubled. Paying for accuracy reduced reliance on AI, but did not remove the effect entirely.
📲SOCIAL MEDIA
🗞️MORE NEWS
China’s cyberspace regulator registered Apple Intelligence for use on iPhones, clearing a required step before launch. Alibaba says Qwen will support the service across Apple devices, while Baidu is helping build features for Chinese iPhone users; no launch date has been announced.
The European Commission ordered Google to give rival AI assistants access to 11 Android functions, including voice-activated searches and bookings, by July 2027. Google must also share qualifying anonymized search data from January 2027, while retaining security checks; Google argues the rules could create privacy and cybersecurity risks.
Nvidia is working with Fujitsu, Fanuc, Yaskawa Electric, and Kawasaki Heavy Industries on robots that can adapt and work beside people. Government-backed Noetra plans to acquire 27,500 Rubin chips, with facilities expected to start operating in June 2028, but the partners have not set firm deployment targets.
Google has delayed Gemini 3.5 Pro because the model has not met internal targets, especially for coding, Bloomberg reported through Reuters. The release is reportedly months behind schedule; Google has not confirmed the details publicly, so the timing and performance claims remain unresolved.
What'd you think of today's edition? |
Reply