ChatGPT Desktop Gets Voice Control, Multi-Agent Collaboration

OpenAI launched ChatGPT Voice in the desktop app, enabling voice control of your computer and directing multiple agents running in ChatGPT Work or Codex — powered by GPT-Live on a unified model for speaking, listening, and coordination simultaneously.

OpenAI announced that ChatGPT Voice is now available in the desktop application, marking a significant expansion of multimodal interaction capabilities. Users can control their computer and orchestrate multiple agents running in ChatGPT Work or Codex entirely through voice commands. The feature is built on GPT-Live, a unified model architecture that simultaneously handles speaking, listening, and task coordination — eliminating the lag and turn-taking friction of earlier voice interfaces. The rollout began globally on July 23, with availability across all desktop platforms. This release positions voice as a primary interface for agent-based workflows, not merely a convenience feature.

Black Forest Labs Unveils FLUX 3: Multimodal Unifying Image, Video, Audio, and Action

FLUX 3 is a frontier multimodal model from Black Forest Labs that jointly learns from image, video, and audio to build a unified world representation. Early access available, expected to be open-sourced.

Black Forest Labs officially released FLUX 3, a multimodal frontier model that jointly learns from image, video, and audio data to build a unified world representation. The model supports text-to-video generation of up to 20 seconds with multiple shots in a single pass, along with image-to-video and video-reference generation capabilities. Early access is now available, and the model is widely expected to be open-sourced, following BFL's established pattern. FLUX 3 represents a significant step toward unified generative models that treat images, video, and audio as first-class modalities within a single architecture, rather than stitching together separate specialized systems.

Andrew Ng Launches OpenWorker: Open-Source Agent That Delivers Work, Not Chat

Andrew Ng announced OpenWorker, an open-source agent that goes beyond conversational interaction to directly deliver completed work. It can prepare customer briefs, draft reports, triage emails, send Slack messages, and update calendar entries — acting as an autonomous task executor rather than a chatbot. The project marks a shift in the agent paradigm from "assistants that talk" to "workers that produce."

FLUX 3 Video: 20-Second Multi-Shot Generation from a Single Prompt

Black Forest Labs also released and open-sourced the video generation capabilities of FLUX 3. The model supports text-to-video, image-to-video, and video-reference generation, producing up to 20 seconds of multi-shot video in a single pass. Detail fidelity is reported to be exceptionally high. The open-source release is expected to energize the video generation ecosystem significantly.

Alibaba Qwen Releases Qwen-Audio-3.0-TTS with Fine-Grained Inline Voice Tags

Alibaba's Qwen team introduced Qwen-Audio-3.0-TTS, available in two versions: Flash for real-time interaction and Plus for high-quality generation. The standout feature is fine-grained inline tag control — users can insert tags like [whisper], [angry], [breaths], and [laughs] directly into text to steer emotional expression and delivery. The model also supports free-style natural language control for nuanced voice synthesis, representing a new level of expressivity in TTS systems.

Grok 4.5 Solves ~30-Year-Old Graph Theory Conjecture

Elon Musk claimed Grok 4.5 has just solved the Graffiti conjecture, a graph theory problem that has remained open for approximately three decades. The claim quickly went viral, attracting over 4 million views.

DeepSeek's Liang Wenfeng to Investors: Open Source Is the Only Agenda, Restraint Is Key to Survival

DeepSeek founder Liang Wenfeng took a forceful stance in a four-hour investor conference call. He insisted that open-sourcing is the sole agenda at DeepSeek and declared bluntly: if you are not on board, go buy another model. More significantly, he argued that restraint — not aggressive expansion — is the only path to survival in the AI industry, warning that "world-eating labs will be culled." The leaked transcript has sparked intense discussion across the AI community.

Chinese 14nm Chip Achieves Hopper+ Level Compute Performance

A 14nm-based Chinese chip has reportedly reached Hopper+ level utility, delivering 6.4 TB/s bandwidth with a roadmap to 20 TB/s by 2027 — equivalent to NVIDIA's Rubin level. While running hot, the chip is seen as a major milestone in China's push for compute independence. Commentators noted that electricity consumption is not a binding constraint in China's context, making this a viable path.

Google DeepMind Launches Gemini 3.5 Flash Cyber for Security Teams

DeepMind introduced Gemini 3.5 Flash Cyber, a specialized lightweight model designed to help cybersecurity teams automatically discover and patch vulnerabilities before they can be exploited. The model is optimized for security workflows, distinguishing it from general-purpose LLMs. The move signals growing interest in domain-specific model variants tailored for operational security use cases.

Ant Group Releases Ling-3.0-flash: 124B MoE with Only 5.1B Active Parameters

Ant Ling released Ling-3.0-flash, a 124B-parameter Mixture of Experts model with only 5.1B active parameters at inference. It features KDA+MLA hybrid attention and supports 256K context length. SGLang is providing day-zero support for production deployment.

Apple-π Benchmark Anchors Video Model Evaluation in Physical Laws

Apple-π is the first benchmark to anchor video model evaluation in physical laws rather than human preference. It includes 400 classical mechanics videos and a three-stage reasoning protocol: perception, formulation, deduction. Among 11 tested models, the best score was only 0.473, revealing a significant perception-to-deduction bottleneck and weakness in multi-law state transitions.

NVIDIA Deploys DGX GB300 at Naval Postgraduate School

NVIDIA CEO Jensen Huang commissioned the DGX GB300 system at the Naval Postgraduate School, providing 1,500 students and 600 faculty with on-premises access to large-scale AI computing. The event, Converge @ NPS, brought together federal leaders and ecosystem partners to mark the activation.

Liang Wenfeng: 4 Ascend 950DTs Equal 1 GB300, Same Latency, Same Capability

Liang Wenfeng stated unequivocally that everything a GB300 can do, a 950DT can do too — at an effective 4:1 exchange rate. One full SuperPod of 8,192 NPUs equals 2,048 GB300s, or roughly 28 NVL72 GB300 racks consuming approximately 3.8 MW. He claims latency and capability are identical. Policy analysts noted Wenfeng dismisses any notion of NVIDIA's software advantage entirely.

vLLM Enables Trillion-Scale Agentic RL Inference

PrimeIntellect's prime-rl 0.6.0 runs on vLLM with FP8, wide expert parallelism, prefill/decode disaggregation, and KV cache offloading. The system trained GLM-5 on SWE tasks at 131k steps, demonstrating trillion-scale agentic reinforcement learning on the inference side.

Ethan Mollick Publishes New AI Guide: Agent Systems Now Extremely Powerful

Ethan Mollick's latest guide for non-experts highlights the shift from chatbots to agentic systems. For high-priority tasks, he recommends Claude or ChatGPT's strongest paid models, which can autonomously complete hours of human-equivalent work.

Plasma AI Open-Sources Fractal: Architecture for Large-Scale Autonomous Coding

Plasma AI's Fractal overcomes the single-context, linear-path limitation of coding agents, enabling autonomous work across multiple services and repositories. It is designed for migrations, audits, and system-wide refactoring tasks.