Claude Opus 4.7, Qwen 3.6, Happy Oyster, realtime 3D worlds, new Google TTS: AI NEWS

TL;DR: Anthropic drops Claude Opus 4.7, Alibaba ships open-source Qwen 3.6 plus a real-time 3-D world generator, Google unveils emotion-rich TTS, and a flood of tiny-yet-powerful open models hit phones, labs, and VR pipelines. ## Prompt Relay: seamless multi-scene video with zero training A plug-and-play trick called **Prompt Relay** layers on top of Alibaba’s Wan model to chain wildly different shots into one smooth clip. Example timeline - 0–2 s: eagle soaring - instant cut to cyber-punk street race - pull-back to living-room TV Instead of bleeding prompts, the method feeds Wan a list of prompts with start/end frames. Inside the cross-attention layers the current prompt dominates while the next one “takes the baton” during a short overlap, keeping motion and style coherent. Code is already posted; full repo drops within days (link in video description). ## Turnary Bonsai: 1.58-bit LLMs that run on your phone The **Turnary Bonsai** family compresses every weight to -1, 0, 1 plus a shared scale, slashing disk size 9× versus 16-bit. Three sizes—1.7 B, 4 B, 8 B—ship as open weights. - 8 B model: 1.7 GB file, outscores Llama-3.1-8B, GLM-4-9B and Ministral-8B on MMLU, BBH, HumanEval - Runs ≥100 tokens/s on consumer GPUs and flagship phone SoCs GitHub + Hugging Face links in description. ## GPT-Rosalind: OpenAI’s life-science reasoning model **GPT-Rosalind** targets the 10-15-year drug-discovery slog. The model chains literature review, hypothesis generation, experiment design and data analysis in one chat. Benchmarks show double-digit gains over GPT-5.4 on molecular-property prediction and protocol-planning tasks. A Code Interpreter plugin wires Rosalind into 50+ bio-databases (UniProt, PDB, PubChem, etc.). Access is invitation-only for wet-lab researchers—apply via the link below. ## WildDebt-3D: iPhone-grade open-set 3-D detection **WildDebt-3D** spits out metric 3-D bounding boxes in real time on an iPhone camera feed. Type “monitor” or “paper” and the network segments, depth-estimates and tracks the object; images work too (“animal” boxes every creature). Weights and Swift demo code are MIT-licensed. ## Motif-Video-2B: tiny video diffusion that punches above its weight A 2-B diffusion transformer trained on <10 k GPU-hours and <10 M clips rivals Alibaba’s 12-B Wan on VBench. Needs only 19 GB VRAM; ComfyUI nodes coming. Hugging Face repo is live. ## Sponsor: HubSpot’s free “AI Content Team” playbook HubSpot’s 11-skill framework shows how to feed top-performing assets into an AI swarm that reverse-engineers viral DNA, spots trending angles, schedules posts and writes fresh copy that improves itself from analytics. Download link in description. ## Annigen: single image → rigged 3-D character Upload one photo and **Annigen** returns a clean mesh, skeleton and skinning weights ready for Maya or Blender. Demo shows desk-lamp, shark and dog assets animated out-of-the-box. Requires 18 GB VRAM; Apache-2.0 code on GitHub. ## Happy Oyster: open-source answer to Google Genie 3 Alibaba’s ATH Lab drops **Happy Oyster**, a real-time, prompt-controllable 3-D world generator. Text like “ride a dragon” or “skateboard over rooftops” spawns an explorable environment in seconds. ATH is also previewing **Happy Horse**, a video model that just beat Seedance 2.0 on VBench. Happy Oyster weights and demo request form are in the description. ## LRA 2: Nvidia’s consistent 3-D scene reconstruction **LRA 2** turns casual video into a persistent 3-D Gaussian-splat scene. Geometry and texture stay frozen—walk away for an hour, come back, everything is exactly where you left it. Code and paper linked below. Source: [YouTube video](https://www.youtube.com/watch?v=G8fqduzB5lc)