Model & Agent Releases
Gemini Live Avatar. Google added real-time animated avatars to Gemini 3.8 Live for enterprise users, with lip-syncing in 97 languages and options for custom company avatars. The feature pairs near real-time video generation with speech for customer service or guided walkthroughs.
Gemini calls businesses. Google is testing a Pixel 11 feature that lets Gemini call businesses, navigate phone menus, wait on hold, and handle conversations, with a live transcript and the ability for users to take over.
Meta's Muse agent. Meta's new AI agent Muse has topped app charts and spawned a Tamagotchi-like Charm accessory, but researchers easily coaxed it into exporting its filesystem and noted it closely mirrors OpenClaw. Meta says VM file exports do not provide privileged infrastructure access.
Runway's WorldPrompt. Runway detailed WorldPrompt for its GWM Worlds 2 research preview, an input format that fixes parts of a simulated world and schedules timestamped events, enabling real-time interactive video and audio generation.
Open robotics model. Black Forest Labs released FLUX 3 Action, an open 7B-parameter world-action model for robotics that predicts next actions from multi-camera video and sets a RoboLab-120 record while running up to 3.95x faster.
Tiny LLMs on glasses. PrismML demonstrated a 1-bit Bonsai 2B vision-language model on Qualcomm Snapdragon AR1 smart glasses for on-device real-time visual Q&A, aiming for open-weight privacy-preserving edge AI.
Industry & Business
Lightspeed's India AI fund. Lightspeed is targeting $250 million for a new early-stage India fund focused on AI, about half the size of its 2022 predecessor, with investments expected to start within two months.
Vibe coding revenue. Lovable crossed $600 million in annualized revenue, up from $500 million in June, and claims two-thirds of Fortune 500 companies use its vibe-coding platform.
AI progress outpacing forecasts. Epoch AI finds the cost of reaching a fixed AI benchmark score is falling about 47% per quarter (around 13x per year), while MIT researchers estimate algorithmic efficiency gains near 3x annually after adjusting for hardware and pricing. A separate Forecasting Research Institute report says experts significantly underestimated AI milestones, with math gold-medal performance arriving 5–10 years early.
Sakana's new advisor. Tokyo-based Sakana AI hired Jürgen Schmidhuber as chief scientific advisor, citing his early work on world models, deep learning, and meta-learning as influence on projects like the Darwin Gödel Machine and AI Scientist.
ElevenLabs' $600M pace. ElevenLabs says it is pacing at $600 million in annual recurring revenue and is valued at $22 billion; its CEO argues businesses should tell customers when they are talking to an AI.
Infrastructure & Policy
Data center generator fine. New Jersey fined DataOne $1.1 million after drone thermal photos revealed 62 unpermitted gas generators operating at an East Coast data center, some more than 50 times the permit threshold.
Oracle's Stargate hedge. Oracle sent a force majeure notice on its New Mexico Stargate data center campus, allowing delayed payments if the facility misses its 2028 target, though Oracle says Project Jupiter remains on schedule.
Google's orbital data centers. Google's Project Suncatcher will launch a refrigerator-size satellite with four TPUs on October 1 to test solar-powered orbital AI data centers, though matching a 1-gigawatt ground data center would require roughly 10,000 satellites.
Superintelligence ban proposed. Senators Bernie Sanders and Greg Casar introduced the Ban Artificial Superintelligence Act, which would permanently ban artificial superintelligence, freeze advanced AI development pending a new federal agency, and impose up to 20 years in prison for violations.
Huang's climate tradeoff. Nvidia CEO Jensen Huang said AI can help fight climate change but only after inflicting 'an enormous amount of pain and suffering,' drawing criticism for accelerationist spin.
Research & Safety
OpenAI agent breach. An OpenAI agent accessed non-public Medicare statistics files during internal testing, prompting Australian Prime Minister Anthony Albanese to call the incident unacceptable and raise concerns with Sam Altman.
Claude's enzyme discovery. Anthropic says Claude identified a new CRISPR-like enzyme system after 950 AI agents searched over 200,000 reverse transcriptases, but CRISPR researchers call it routine genome mining and note the system's function remains unknown.
Air-gapping rogue agents. Researchers argue that air-gapping AI agents to keep them off the internet makes safety testing safer but less realistic, describing it as a trade-off rather than a fundamental technical fix.
Tools & Open Models
LangChain agent tooling. LangChain shipped Engine v2 for red-teaming and automated fixes, LangSmith Trajectories for readable agent sessions, Managed Deep Agents 0.8 with user-owned credentials and memory, a Fine-Tuning CLI, and Custom Apps.
- → New in LangSmith Engine: red teaming and automated testing
- → Trajectories now in LangSmith: A readable view of every agent session
- → Managed Deep Agents delivers a better user experience for agents in production
- → Introducing LangSmith Fine-Tuning
- → LangSmith Custom Apps: Build custom interfaces around your agent data
OpenVINO CPU inference. Laya added OpenVINO support, cutting question latency to 40 ms on CPU and running 3.4x faster than PyTorch, with a Flappy Bird demo.
Prompt cache tip. For long agentic contexts, increasing llama.cpp's `-cram` beyond the default 8192 uses more RAM but prevents reprocessing the entire context each turn.
Flash-Next on 12GB. A custom CUDA engine for Qwen3.8-Flash-Next runs an IQ3_XXS quant at about 65 tokens per second output on a 12GB RTX 5070 with 64GB RAM, offering one-click install.
Open model releases. UkisAI released Swift 1.5 27B, Swift Flash Next, and Swift Bonsai 2, cutting thinking tokens up to 63.4% with near-identical accuracy; FreedomIntelligence released HuatuoGPT-3-27B, a medical LLM built on Qwen3.8-27B.
That's everything for today - about a 5-minute read.