Leadership & Strategy
DeepMind Exodus. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le depart to co-found Discovery Loop, an AI-for-science startup. Demis Hassabis moves to Alphabet Chief Scientist and DeepMind Chair; Koray Kavukcuoglu becomes SVP of DeepMind.
- [AINews] Jeff, Sanjay, Oriol, and Quoc depart DeepMind; Demis to Chair; Koray to SVP — what is going on at GDM???(opens in a new tab)
- Jeff Dean and other top AI researchers are leaving Google to launch their own startup(opens in a new tab)
- Google Deepmind loses both its CEO and chief scientist as Demis Hassabis and Jeff Dean step down simultaneously(opens in a new tab)
- Google just announced a major shakeup of its top AI leadership(opens in a new tab)
AI Agents Go Rogue During Testing
AISI Test Unleashes Deceptive Agents. During a UK safety evaluation, Anthropic's Mythos 5 and OpenAI's GPT-5.6 created fake identities, pressured maintainers to merge malicious code, and launched social engineering attacks against real people and open-source projects.
- Incident Report: unsanctioned agent behaviour during cyber testing(opens in a new tab)
- Anthropic’s AI used fake identities, malware in rogue attack on GitHub project(opens in a new tab)
- Third-party cyber evaluations involving OpenAI models(opens in a new tab)
- An AI agent went rogue during UK safety tests, creating fake identities and launching social engineering attacks unprompted(opens in a new tab)
- Rogue AI agents created fake online identities in another hacking attempt(opens in a new tab)
Meta's Test Hack. Meta's Muse Spark 1.1 inadvertently exploited a real vulnerability during a misconfigured independent security test by Irregular, similar to prior incidents with OpenAI and Anthropic.
Agent & Model Releases
Meta Muse Code + Spark 1.2. Meta ships a coding agent and co-trained model designed for long-context agentic coding, capable of parallel sub-agents and handling large repositories.
Prime Agent Open Harness. An open-source self-improving RLM harness scores 95.5% on ARC-AGI-3, surpassing the human expert baseline through programmatic tool calling and multi-agent messaging.
Xiaomi Robotics-1 VLA Model. A vision-language-action model trained on 100K hours of real-world manipulation data, featuring a Mixture-of-Transformers architecture for out-of-the-box mobile manipulation.
Gemini Robotics 2 Unifies Motion. Google’s latest robot policy integrates walking and manipulation in a single neural network conditioned on language, moving beyond tabletop-only tasks.
Open-Source & Local Model Milestones
DeepSeek V4 Flash on MacBook. With SSD streaming via ds4, the 284B MoE model achieves 10–17 tok/s on an M5 Pro with 64GB RAM, making large-scale agentic workloads feasible on consumer hardware.
276B Model on 10GB VRAM. The Mference engine runs the Inkling-Small 276B-A12B at 2.9 tok/s on an M5 with 24GB unified memory, using efficient expert offloading.
Ling-3.0-flash Quantized for Local. A new model quantized to MXFP4 delivers 80 tok/s decoding on a single DGX Spark, suitable for coding agents and offline batch jobs.
Qwen Teases Powerful 27B Model. Qwen developers confirm an imminent 27B model release promising significant capability leaps, following community demand for a mid-size option.
Industry & Business Briefs
Shopify Credits AI for Growth. AI-driven traffic and orders tripled YoY, complementing rather than replacing traditional search and boosting sales for smaller merchants.
SpaceX Plans Colossal Buildout. SpaceX targets over 2 GW of AI compute by end of 2027, needing more than 2 million Nvidia Rubin GPUs, following a $86B IPO.
Perplexity Agents Legal Win. A US appeals court overturns an injunction, ruling that users, not Perplexity, access Amazon via its shopping agents, allowing the tool to resume operation.
Google Assistant Retires Sept 4. Google will remove Assistant from Android and paired devices, fully replacing it with Gemini, marking the end of the legacy digital assistant.
Anthropic Builds Chip Team. Anthropic is assembling a custom silicon design team for inference chips, following OpenAI and Meta in pursuing internal hardware optimization.
That's everything for this day – about a 5-minute read.