Saturday, August 15: AI builds, protects, and labels

One verified August 15 launch leads this episode, followed by two notable August 14 drops. Together, they show AI moving from generating an output to clarifying the task, protecting the input, and preserving provenance.

SoonLab announced SoonLab 2.0 as an AI-native game-creation model. The company says it expands natural-language creation from 2D into playable 3D games, while an agent asks follow-up questions about mechanics, rules, difficulty, and visual behavior before building. It is available through SoonLab’s site. The announcement was carried by Yahoo Finance as a paid press release, so its capability claims are best treated as the maker’s launch description rather than an independent test. 1

Google introduced HEIR, short for Homomorphic Encryption Intermediate Representation, as an open-source compiler and toolchain for private AI inference. It is designed to help convert models that operate on unencrypted inputs to operate on encrypted data, so a service can compute on information without seeing the underlying content. Google also published four example applications and source code. The project is a developer tool and research platform, not a turnkey private-AI service; Google’s stated goal of making encrypted inference easier for non-experts remains a direction, not a guarantee that every deployment is one click. 2

Google is also changing how generated media is presented. TechCrunch reports that Gemini and Flow are rolling out a Media Watermark setting that lets users turn off the visible mark on images, videos, and songs made with Nano Banana, Omni, and Lyria. The invisible SynthID watermark and C2PA metadata remain, and Search support is still coming. The setting is therefore a presentation control, not a removal of provenance; rollout timing and surface availability still matter. 3

The useful question for a closer look is which boundary matters to you: a playable prototype, privacy-preserving inference, or control over how generated media appears. None of these launches proves the whole workflow is solved, but each makes one part of the work layer more explicit.