GPT 5.6, Mythos ban lifted, realtime avatars, Seedance 2.5, brain ultrasound: AI NEWS

The video surveys a wave of rapid breakthroughs across AI video, hardware, and generative modeling, highlighting ByteDance’s Seedance 2.5 and Alibaba’s lifelike real-time interactive avatars as major steps toward human-like digital interaction. It also covers new silicon, including OpenAI’s Broadcom-built “Jalapeno” AI chip and IBM’s sub-1-nanometer transistor breakthrough, alongside powerful but restricted models like GPT-5.6 and Claude Mythos. Additional highlights include Domain Shuttle’s reference-based video generation that preserves characters across styles, Ornith’s open-source agentic coding models that outperform much larger systems, and Stability AI’s Arbor system for geometry-constrained 3D generation. The video closes with ByteDance’s Dance OPD framework, which unifies text-to-image, editing, and style transfer into a single model by resolving conflicting training guidance. Overall, the message is that AI is rapidly converging on multimodal, interactive, and physically controllable systems—though many advances remain early-stage, resource-heavy, or limited in public access.

• [0:04-0:18] ByteDance's Seedance 2.5 video model and Alibaba's real-time interactive avatars highlight a major push in AI video generation.
• [0:18-0:29] Hardware breakthroughs include OpenAI's AI-specialized chip and IBM's "revolutionary" subnanometer chip.
• [0:40-0:54] Powerful new models—OpenAI's GPT-5.6 and the un-banned Claude Mythos—are unveiled but likely restricted for most users.
• ▶ 1:20 Live demos show natural back-and-forth conversation: the avatar handles riddles, acknowledges corrections gracefully, and responds to emotional cues like tiredness.
• ▶ 2:09 The avatar's realistic facial expressions, gestures, and casual body language make interactions feel human-like and believable.
• ▶ 3:00 One Streamer uses a single transformer model interleaving text, audio, and video, running at 25 FPS with ~200 ms latency and full-duplex listening/watching while speaking.
• ▶ 3:32 It's still early-stage (v0.1): low-resolution 192p demos, no public code—only a technical paper and main page link are available.
• ▶ 3:48 Domain Shuttle is a reference-based video generation model that takes reference images and generates videos keeping those subjects visually consistent throughout.
• ▶ 4:02 It supports multiple characters and objects as inputs, and can transform them into different styles (e.g., anime) via text prompts.
• ▶ 4:20 The model handles complex scene prompts, such as characters interacting with holographic or 2D graffiti versions of other referenced subjects, with use cases including influencer and product videos.
• ▶ 4:44 The domain shuttle model can generate UGC-style videos and commercials from reference images.
• ▶ 5:14 It excels at preserving multiple reference characters even when they come from different art styles, which competitors fail to do.
• ▶ 5:46 Code is publicly released, but the 35GB model requires a high-end GPU to run locally.
• ▶ 6:15 Ornith 1.0 is a new open-source family of mixture-of-experts models for agentic coding, ranging from 9B to 397B parameters.
• ▶ 6:40 It reportedly outperforms much larger open models like GLM and DeepSeek V4 (1.6T) on agentic benchmarks, with the 35B version beating Qwen 3.5/3.6 and Gemma 4.
• ▶ 7:31 Key innovation: Ornith learns to generate its own harness/scaffold, improving its strategy, memory, error handling, and tool use autonomously.
• ▶ 8:29 Stability AI, the team behind Stable Diffusion, announces a new 3D generation system called Arbor.
• ▶ 8:36 Arbor moves beyond text-prompt-only 3D generation by letting users directly specify where geometry should and shouldn't exist.
• ▶ 8:44 This enables spatial control over the output, marking a shift toward more direct, geometry-constrained 3D generation.
• ▶ 8:48 Arbor uses spatial constraint meshes—simple “rough spatial blueprints” rather than finished 3D models—to let users specify where object parts contact or fit.
• ▶ 8:56 Users can define practical constraints, such as a chair fitting in a volume, leaving empty space under the seat, and having legs touch the floor at specific spots.
• ▶ 9:14 These constraint meshes are fed directly into an existing 3D generation model without retraining, and tests show Arbor obeys spatial requirements far better than previous approaches.
• ▶ 9:42 The code for the featured system has already been released, with a "Code" button on the project page providing full local setup instructions.
• ▶ 9:56 The model is "fairly tiny" at roughly 600 MB, but it also requires Trellis and Trellis 2 3D model generators.
• ▶ 10:07 The system can be used directly as a Blender add-on, with setup instructions included on the page.
• ▶ 10:18 OpenAI is reportedly making "Jalapeno" chips in partnership with Broadcom, described as their first intelligence processor.
• ▶ 10:29 The chip is designed specifically for running AI models, handling workloads that power products like ChatGPT and Codex.
• ▶ 10:33 OpenAI and Broadcom's "Jalapeno" AI processor targets ChatGPT/Codex workloads with faster, cheaper, and more efficient performance; early tests show substantially better performance per watt.
• ▶ 11:39 IBM announced the first sub-1-nanometer chip technology (0.7nm / 7-angstrom node), using a new 3D "nano stack" architecture that vertically stacks and staggers transistors to overcome physical scaling limits.
• ▶ 12:27 IBM's sub-1nm technology fits nearly 100 billion transistors on a fingernail-sized chip (~twice the density of its 2nm chip) and delivers up to 50% more performance and 70% better energy efficiency.
• ▶ 13:10 A major challenge in AI image generation is that creating a single model that handles text-to-image, image editing, and style transfer is very hard; current top models are often text-to-image only, requiring separate models for editing.
• ▶ 13:39 ByteDance's Dance OPD framework aims to solve this by building one unified model that is great at generating images from text, editing specific parts, and applying artistic styles.
• ▶ 13:59 The key method: instead of blending multiple specialist models, Dance OPD treats each capability as a separate field, letting the student model learn from each "teacher" individually to avoid conflicting guidance during training.
• ▶ 14:19 The Dance OPD training framework resolves conflicting instructions by selecting a single teacher model for each training step; the student model asks that teacher what to do and learns to match its behavior.
• ▶ 14:32 Using this framework, a single model is trained to handle multiple tasks: text-to-image, image-to-image editing, and style transfer.
• ▶ 14:46 Quantitative results are “really good,” with Dance OPD scoring the highest when combining text-to-image and editing capabilities, outperforming other methods.
• ▶ 14:53 Method achieved top benchmark scores, outperforming competing approaches.
• ▶ 14:57 A technical paper documenting the work has been officially released.
• ▶ 15:00 The accompanying code is coming soon, to be made publicly available.
• ▶ 15:14 Higsfield now enables 4K video generation with Seed Dance 2.0, a major upgrade from the previous 1080p limit.
• ▶ 15:33 The high resolution means creators can crop, resize, add text, and repost across platforms without blurriness or quality loss.
• ▶ 15:54 Beyond resolution, the update delivers consistent characters across shots, multi-shot scenes, real camera movement, and multimodal input mixing text, images, video, and audio.
• ▶ 16:49 Unzero, a new image generation architecture by Unconventional AI, is introduced as a fundamentally different alternative to diffusion models.
• ▶ 17:14 Instead of denoising random noise, Unzero generates images by coupling thousands of oscillators that synchronize into a coherent picture, potentially avoiding slow iterative steps.
• ▶ 18:08 The project is fully open source under the MIT license, with model weights and training scripts available, though current results are only a proof of concept rather than production-ready.
• ▶ 18:42 Humanoid robots need to learn from real-home data, not just lab demos, covering chores like setting tables, laundry, sweeping, and organizing objects.
• ▶ 19:32 HIW500 includes over 500 hours and 10 TB of whole-body teleoperation data from the Uni31 robot in actual homes, spanning over 10 tasks with varied lighting, objects, and clutter.
• ▶ 20:01 The dataset is open-sourced: the raw data (~9 TB) is available via a linked resource, and it will also be released in Hugging Face's LeRobot format for robot training.
• ▶ 20:27 ByteDance unveiled its latest frontier video generation model, Seed Dance 2.5.
• ▶ 20:33 The previous version, Seedance 2.0 O, remains by far the leading video model, with no other model coming close to beating it.
• ▶ 20:45 Google released Gemini Omni a few months later, but it still did not surpass ByteDance's model.
• ▶ 20:52 Seedance 2.5 is a major upgrade to ByteDance's video model, adding longer 30-second clips, up to 50 reference images, precise box-based local editing, stronger character/scene consistency, native audio, and up to 4K resolution.
• ▶ 21:53 The biggest improvement is control: creators can input many reference images and edit existing videos with bounding boxes, making it a production tool for filmmakers and branded content rather than simple text-to-video.
• ▶ 22:21 The model is set to launch in early July, with the presenter planning deeper coverage later.
• ▶ 22:40 Seed 2.1 is designed specifically as a productivity agent, with Pro (more performant) and Turbo (faster) variants.
• ▶ 22:54 On real-world benchmarks, it is on par with Claude Opus 4.7 and GPT 5.5, and roughly on par in Agentic Coding.
• ▶ 23:11 Its standout strength is multimodal understanding, outperforming Opus 4.7 and GPT 5.5 on chart/graph analysis and beating the latest Gemini models on video/motion analysis.
• ▶ 24:32 Lyft 4D reconstructs a full 4D scene (3D + time) from a single ordinary 2D video, including occluded or unseen regions, enabling orbiting around dynamic objects.
• ▶ 25:38 ByteDance's Perception DM uses a diffusion language model to caption multiple image regions simultaneously, outperforming similar-size vision models and is fully open-sourced.
• ▶ 27:01 Alibaba's Happy Horse 1.1 video model offers 1080p, 3–15 second clips, native audio with lip-sync, better motion/character consistency, but still trails Seedance.
• ▶ 30:27 OpenAI released GPT 5.6 (Soul/Terra/Luna family), but almost no one can access it yet—only a limited preview for trusted partners due to US government compliance, raising concerns about elite-only access.
• ▶ 31:34 GPT 5.6 Soul Ultra sets a new state-of-the-art on Terminal Bench and beats Claude Mythos, though it still trails Mythos on cybersecurity exploit benchmarks; OpenAI used 700,000+ GPU hours of red teaming for safety.
• ▶ 33:29 OpenAI may postpone its IPO to next year, likely tied to the new US frontier-model ban—and the government is easing bans on Claude Mythos 5 only for ~100 trusted partners, reinforcing the case for open-source AI like GLM, DeepSeek, Qwen, and MiniMax.
• ▶ 35:54 Olive unveiled an ultrasound-based brain imaging method that fills the gap between invasive electrodes and blurry external EEG, using ultrasound to track blood flow and map brain vessels in high detail without drilling into the skull.
• ▶ 37:11 The technique uses injected microbubbles to achieve 100x greater resolution than similar CT scans, though it is not fully non-invasive since the contrast agent must enter the bloodstream; Olive is also open-sourcing the entire pipeline and dataset.
• ▶ 38:38 Meta's Auto Data framework autonomously builds high-quality training/evaluation data via a self-improving loop of generation, challenge-based evaluation, and performance measurement, showing that a small model trained on its output outperformed a standard method.
• ▶ 40:21 Sakana released "Fugu" models claiming they surpass frontier models, but the presenter calls this misleading because Fugu is not a standalone base model.
• ▶ 40:38 Fugu is actually an orchestrator/router that combines multiple closed and open models, making its benchmark comparisons an apples-to-oranges ensemble method.
• ▶ 41:17 The service is paid, closed-source, and expensive with no transparency about which models are used, leading the presenter to call the situation "very fishy."

• ▶ 0:00 AI Video, Coding, and 3D Generation News (0:00 - 10:18) - Weekly AI recap, One Streamer, Domain Shuttle video ads, Ornith 1.0 coding agent, and Stability AI's Arbor 3D generation.
• ▶ 10:18 AI Hardware, Dance OPD, and Sponsored Segment (10:18 - 16:49) - OpenAI-Broadcom Jalapeno, IBM Nano Stack, Dance OPD framework, and sponsored Higsfield/Seed Dance 2.0 4K.
• ▶ 16:49 Image Generation, Robotics Dataset, and Bite Dance Models (16:49 - 24:32) - Unzero's coupled-oscillator image model, HIW500 humanoid dataset, Seed Dance 2.5 video model, and Seed 2.1 family.
• ▶ 24:32 4D Reconstruction, Video/Image Models, and Humanoid Robots (24:32 - 30:27) - Lyft/Lift 4D reconstruction, Perception DM captioning, Happy Horse 1.1, Korea 2 image generator, and Uni Tree R1 robot demo.
• ▶ 30:27 OpenAI GPT 5.6, IPO, and Claude Mythos Controversy (30:27 - 35:54) - GPT 5.6 release, features, speed limits, OpenAI IPO reports, and Anthropic's Mythos/Claude open-source debate.
• ▶ 35:54 Brain Imaging, Open-Source Pipeline, and Meta Auto Data (35:54 - 40:21) - Olive's ultrasound brain imaging, Alzheimer's/TBI pipeline, and Meta Auto Data.
• ▶ 40:21 Sakana Fugu and Video Wrap-Up (40:21 - 42:44) - Sakana Fugu models and closing call to action.