Agentic AI News Sept. 24, 2026

Meta Muse anywhere. Meta is expanding Muse with its own email addresses, computer use from the Mac app, and live video calls with customizable avatars, alongside support for Meta smart glasses.

• Everything new coming to Meta’s AI agent Muse(opens in a new tab)
• Muse is coming to Meta smart glasses(opens in a new tab)
• Meta is making Muse more powerful and will let you video chat with it, too(opens in a new tab)

Muse Charm and audio glasses. Meta revealed the Muse Charm, a palm-sized standalone wearable with the Muse avatar and fingerprint sensor, plus camera-free Ray-Ban Meta Audio glasses starting at $349.

• Meta made a Tamagotchi-like wearable for its Muse AI agent(opens in a new tab)
• Meta is making a standalone Muse AI gadget(opens in a new tab)
• Meta introduces camera-free AI glasses(opens in a new tab)
• Meta ditches the camera on its newest smart glasses(opens in a new tab)

Muse adoption and OpenClaw. Muse drew over 500,000 users in its first week, and Meta admitted it is heavily inspired by the open-source OpenClaw project; a hands-on review found it capable but eager to spend money.

• Meta's AI agent Muse draws 500,000 users in a week along with claims it copied OpenClaw(opens in a new tab)
• Meta’s AI agent is a cute little guy who’s great at spending my money(opens in a new tab)

ChatGPT Voice agentic. ChatGPT Voice now runs on GPT-6 Astra, Sol, and Luna models and can access email, calendar, Slack, and finance apps, with voice control in the mobile Work tab.

• ChatGPT mobile app gets voice-based agentic features(opens in a new tab)
• ChatGPT Voice gets closer to "Her" with email, calendar, and Slack access(opens in a new tab)

YouTube AI tools. YouTube introduced custom AI-curated feeds, Gemini editing tools in Creator Studio, dynamic thumbnails, and conversational features in YouTube Music.

• YouTube promises custom feeds and a lot more AI later this year(opens in a new tab)
• YouTube Music gets more conversational with new AI features(opens in a new tab)
• YouTube will let you build your own algorithm with AI(opens in a new tab)
• YouTube releases new AI features for creators within its Studio app(opens in a new tab)
• YouTube adds AI tools to Creator Studio with script coaching, smart thumbnails, and Gemini editing(opens in a new tab)

Gemini 3.8 TTS. Google's new Gemini TTS models let users design voices from text descriptions or 30-second samples and direct multi-character dialogue across more than 100 languages.

• Gemini 3.8 text-to-speech says hello(opens in a new tab)
• Gemini 3.8 TTS Playground(opens in a new tab)
• Google's new Flash TTS models let you design AI voices from scratch using text descriptions(opens in a new tab)

Apple LensVLM-9B. Apple released LensVLM-9B, a vision-language model that scans compressed images of text and selectively expands relevant pages using learned tools.

• apple/LensVLM-9B · Hugging Face(opens in a new tab)

FLUX 3 Action. BFL released FLUX 3 Action, a 7B robot model available on bfl.ai.

• BFL releases FLUX 3 Action: a 7B robot model(opens in a new tab)

Nemotron 3 Diarization. NVIDIA's Nemotron 3 Diarization model streams speaker labels for up to eight speakers, integrates with Transformers, and supports local voice agents.

• Streaming Nemotron 3 Diarization(opens in a new tab)

MiMo V2.6 and V3. Xiaomi's MiMo-V2.6 Pro and Flash impressed in coding tasks despite questionable benchmarks, and the upcoming MiMo-V3 will use a new HySparse2 architecture.

• MiMo-V2.6 (both Pro and Flash) is a benchmaxxed scam(opens in a new tab)
• MiMo-V3 is getting a new architecture. The core of it, HySparse2, is out today.(opens in a new tab)

Local inference gains. Community posts detail running Qwen 3.8 Flash Next at 15-20 t/s on 16GB VRAM via MTP and CPU offload, plus new high-quality Qwen3.8 27B quants; NVFP4 quality comparisons remain debated.

• Qwen 3.8 Flash Next q4_k_m, 130k context, q8 cache on 16GB VRAM ann 64GB RAM, 15-20 t/s on 4080(opens in a new tab)
• NVFP4 - what does the quality usually equate to? (5090)(opens in a new tab)
• Perhaps the highest quality mainline quants of Qwen3.8 27B?(opens in a new tab)

Graphify code context. Graphify converts codebases into queryable multimodal knowledge graphs to give AI coding agents better cross-file awareness.

• Graphify: Unifying Codebase Context to Streamline Agentic Software Engineering(opens in a new tab)

Modal million sandboxes. Modal rebuilt its infrastructure to support 1 million concurrent sandboxes and tens of thousands of creations per second, bypassing Kubernetes central coordination limits.

• Beyond Kubernetes at Modal: How to Scale 1 Million Concurrent Sandboxes in Seconds(opens in a new tab)

Warp and MjWarp. NVIDIA Warp and MjWarp enable MuJoCo physics models to scale to 2,048 parallel GPU environments for robotics simulation and learning.

• How to Use NVIDIA Warp and MjWarp to Accelerate Robotics Simulation and Learning Workflows(opens in a new tab)

Offloaded robot inference. Microsoft Research shows offloading physical AI inference from onboard GPUs to edge or cloud can improve task success, model size, and battery life, with new Kubernetes-based tooling.

• Offloaded inference for real-world physical AI robotics(opens in a new tab)

Anthropic biolab discovery. Anthropic's wet lab used nearly 1,000 Claude agents to find a previously unknown enzyme system in bacteriophages with properties reminiscent of CRISPR, consuming 210 million tokens over 21 hours.

• Anthropic says its biology lab has already found something big(opens in a new tab)
• Anthropic’s biolab made a discovery it’s comparing to Crispr(opens in a new tab)

Enveda drug discovery. Enveda raised $311 million Series E at a $2 billion valuation to advance AI-discovered nature-derived drug candidates into clinical trials.

• Enveda secures $311M to bring more nature-derived AI drugs into clinical trials(opens in a new tab)

XPRIZE Wildfire AI. XPRIZE Wildfire finalists demonstrated space-based fire detection within 10 minutes and autonomous firefighting drones, though no grand prize was awarded.

• XPRIZE Wildfire winners spotted fires within 10 min—but couldn’t stop them(opens in a new tab)

Claude writing tradeoff. Anthropic's Jackson Kernion says Claude's writing got worse because newer models are optimized for math and code and trained to explain to other AI models, creating dense 'Claudeish' style.

• Anthropic engineer explains why Claude's writing got worse although the model got smarter(opens in a new tab)

US-China AI safety. The US and China began AI safety talks, with Treasury proposing a notification mechanism, but export controls and trust gaps limit cooperation.

• Trump’s China rivalry and “AI race” delusion may endanger US, experts say(opens in a new tab)

California data center disclosure. California signed bills requiring data centers to disclose electricity and water use, giving communities more insight into AI infrastructure impacts.

• Data centers are black boxes, but California wants to change that(opens in a new tab)

Superintelligence ban bill. Bernie Sanders and Greg Casar introduced legislation to ban artificial superintelligence development, with up to 20 years in prison for violations.

• Bernie Sanders proposes banning ‘superintelligence’ and putting violators in prison(opens in a new tab)

Nscale IPO omits ByteDance. Nscale's IPO filing does not name ByteDance, its largest customer accounting for 73% of 2025 revenue, due to legal and reputational risks around US chip export rules.

• Nvidia-backed Nscale keeps its biggest customer, Bytedance, out of its IPO filing(opens in a new tab)

AI worry survey. A Gallup/Microsoft survey found 68% of daily AI users in the US are worried about AI, complicating adoption as a measure of enthusiasm.

• Even Americans who use AI every day are worried about it(opens in a new tab)