Google Vids Adds AI Avatar Feature — Star in Videos with Your Own Digital Twin

On July 16, Google announced that it has added a personalized AI avatar feature to Vids, its video creation tool. Users can create a digital version of themselves and have it appear in their videos. The feature is built on Gemini Omni, and it ships alongside a tool that generates and edits videos from prompts and reference images.

Vids, which had previously been limited to creating presentation materials, is signaling a shift toward becoming a more full-fledged video production tool. While AI is democratizing bodily representation, attention is also focused on how the company will address deepfake concerns.

Roblox Launches AI Game Creation in Its Mobile App

On the same day, Roblox added "Build," an AI-powered game creation feature, to its mobile app. Reportedly, a single text prompt is enough to generate a basic game.

For Roblox, whose creation tools have centered on the PC, the barrier to entry from mobile drops dramatically. The significance of an experience where anyone from kids to adults can build a game with no programming knowledge, entirely on a smartphone, is hard to overstate.

54% of Enterprises Have Experienced AI Agent Security Incidents

A VentureBeat survey of 107 companies has thrown the security problems of AI agents into sharp relief. More than half of enterprises have already experienced a security incident or near miss related to AI agents.

Even more concerning, only about a third of companies assign unique credentials to each agent, with many agents sharing credentials. Only 3 in 10 companies isolate their highest-risk agents. Much of the security stack is repurposed from model providers and hyperscalers rather than purpose-built for agents.

An "agent security gap" is now clearly visible, with the pace of AI agent adoption outstripping identity management and isolation/enforcement controls.

NotebookLM Rebranded as "Gemini Notebook," with Search Opening Up to Apps

Google has renamed NotebookLM to "Gemini Notebook," integrating it more deeply into the Gemini ecosystem. Each notebook is assigned a cloud computer, making it possible to write and run code (initially available to AI Ultra and Workspace customers).

In addition, Google Search is gaining integrations with third-party apps, allowing users to tap into external app features directly from search results. Folding the standalone NotebookLM brand into the Gemini brand suggests that Google's AI product strategy is entering a brand-consolidation phase.

Bloomberg Feature: AI's "Reckoning" Has Arrived for Software Engineers

On July 16, Bloomberg published a major feature on how tools from Anthropic and OpenAI are transforming coding as a profession. The analysis — that the explosive adoption of AI coding tools is driving fundamental change in the very structure of the engineering occupation — takes readers to the front lines of Silicon Valley.

Over the past few weeks, AI acceptance on the ground in engineering has accelerated rapidly, from Bun's Rust rewrite to growing AI tolerance in the Linux kernel. Bloomberg's feature can be read as an attempt to reframe these individual cases within the context of a larger structural transformation.

DFlash: 2.2x Speedup for Qwen3.6 27B with No Quality Degradation

The "DFlash" method, which speeds up Qwen3.6-27B inference by 2.2x with no quality degradation, is drawing attention in the LocalLLaMA community. In real-world measurements on an RTX 6000, the baseline of 44 tok/s improved to 98 tok/s (2.2x) with DFlash applied.

Notably, it recorded 152 tok/s (3.4x) on structured data such as JSON. DFlash generates 15-token drafts at a time, achieving high acceptance rates on repetitive and structured tasks. Meanwhile, for creative tasks the conventional method (MTP) retains the advantage, illustrating that the right choice depends on the use case.

---