01
Odyssey ships Agora-2 multiplayer world model
Odyssey's Agora-2 lets up to 20 humans and agents share one simulated world in real time, and the multiplayer research preview is live to try today.
What they showed / shipped
- Odyssey introduces Agora-2, a next-gen multi-agent world model: up to 20 humans and agents interacting in one shared environment, simulated in real time (@odysseyml).
- Multiplayer research preview is available to try now; teaser posted the day before for an 8:30am PT drop (@odysseyml).
- 📌 On the radar: Min Choi demos ACTx486, where you talk to a video, change the scene, switch languages, and enter the world (@minchoi).
Why it matters
- Shared-world multi-agent sims are the missing testbed between single-agent demos and real multiplayer products.
Sources
02
NVIDIA and DeepMind open 2,800 viral protein structures
NVIDIA, Google DeepMind, EMBL-EBI and partners release AI-predicted protein complex structures for 2,800+ viruses so outbreak prep starts from open data, not a blank page.
What they showed / shipped
- NVIDIA: AI-predicted protein complex structures for 2,800+ viruses are now openly available with Google DeepMind, EMBL-EBI, and research partners (@nvidia).
- Framing is outbreak readiness: give scientists a structural head start before the next viral threat, not after.
Why it matters
- Open structural biology datasets are reusable for docking, vaccine design, and agentic lab workflows.
Sources
03
Google Suncatcher puts ML infrastructure in orbit
Google's Project Suncatcher aims to run ML compute in space; Verge and Ars say the first orbital data-center test launches October 1.
What they showed / shipped
- Google Research facts page: Project Suncatcher puts ML infrastructure in space (blog.google).
- The Verge: Google is sending an AI satellite into space next week (verge).
- Ars: first Suncatcher orbital data-center test launches October 1 (arstechnica).
Why it matters
- If training/inference can leave Earth power and cooling limits, the infra roadmap rewrites.
Sources
04
Midjourney ships edit model, tile, and realtime preview
Midjourney drops a how-to for its new edit model (consistent characters), refreshes --tile and selective inpainting, and previews real-time models on alpha.
What they showed / shipped
- Midjourney posts a video walkthrough of the new edit model for consistent characters and more (@midjourney).
- Same-day updates: new --tile for seamless patterns in v8.1/v8.2, inpainting that only changes selected pixels, and an early look at real-time models on the alpha site (@midjourney).
- 📌 On the radar: Runway adds Draft mode for Seedance 2.5 (faster, fewer credits, then enhance) (@runwayml); Venturetwins shares a Seedance character-swap recipe with reference video + images (@venturetwins).
Why it matters
- Selective inpaint + realtime preview are the primitives that turn image gen into an editor, not a slot machine.
Sources
05
Gemini gets a Live Avatar and can call businesses
Google gives Gemini 3.8 Live a face via Live Avatar, and starts letting Gemini place real phone calls to businesses so you do not sit on hold.
What they showed / shipped
- The Verge: Gemini 3.8 Live with Live Avatar gives Google's AI a face (verge).
- TechCrunch + Verge: Google tests Gemini making phone calls for you, initially for US Pixel owners (techcrunch, verge).
Why it matters
- Voice agents leave the chat box and enter the PSTN; hold-time automation is a real product surface.
Sources
06
Perplexity Computer goes local on AMD plus Fast Search
Perplexity ships Portable Computer for Windows on AMD Ryzen AI Max, and launches Photon Fast Search in the Search API at sub-250ms p95, built with a small team plus hundreds of agents.
What they showed / shipped
- Portable Computer for Windows is live on AMD Ryzen AI Max: local AI agents over your apps and files, including scheduled on-device work (@perplexity_ai, @AravSrinivas).
- Fast Search in the Perplexity Search API runs on Photon, a Rust retrieval/ranking service: 95% of results in 230ms or less, sub-250ms p95 (@perplexity_ai, @AravSrinivas).
- Nous Portal subscribers get Fast Search as default search on Hermes Agent (@AravSrinivas).
Why it matters
- Local agent runtime on Windows NPU silicon plus a public sub-250ms search API is immediately usable infra.
Sources
07
PrismML puts tiny LLMs on Qualcomm smart glasses
PrismML is bringing its tiny LLMs to Qualcomm-powered smart glasses, another bet that on-device models are how AI leaves the phone.
What they showed / shipped
- TechCrunch: PrismML brings its tiny LLMs to Qualcomm-powered smart glasses (techcrunch).
- 📌 On the radar: Meta Horizon Create Studio will let people build AI games on a phone (verge); Meta still drawing heat after a critical AI Glasses video filmed on campus was taken down (reddit).
Why it matters
- Tiny on-device models plus Qualcomm glass SKUs are the practical path for always-on AR agents.
Sources
08
Gen Alpha's insult is "that's so AI" as tutors shut down
The Guardian says Gen Alpha's biggest insult is calling something "so AI," while an Australian tutoring company tells parents to save their money and use AI instead.
What they showed / shipped
- Guardian: "That's so AI" is what Gen Alpha's biggest insult tells us about how kids now police authenticity (guardian).
- AFR + Reddit: an Australian school tutoring company shuts down and tells parents to use AI instead (afr, reddit).
- 📌 On the radar: a poll circulating on Reddit puts China #1 for AI optimism while the US sits near the bottom (reddit).
Why it matters
- If kids treat AI-looking work as socially radioactive, product UX has to look human and earned, not glossy.
Sources