AI finally sleeps, Gemini 4 Carbon, ETH cracked, energy-based models, Anthropic IPO: AI NEWS

Summary

The video is a broad AI-news roundup covering scientific research, new models and tools, open-source releases, and an anticipated Anthropic IPO. The presenter focuses especially on practical capabilities, benchmark claims, hardware requirements, and whether projects can be run locally.

Major research and model announcements

• OpenAI’s claimed scientific results: The presenter describes three purported breakthroughs:
  • An algorithmic result challenging the Exponential Time Hypothesis (ETH), a longstanding assumption about the difficulty of SAT problems. The video says the proof was checked step by step using Lean, while cautioning that the algorithm may still be too slow for practical use.
  • A proof concerning the local smoothing hypothesis, reportedly extending results to any dimension and relating to the Kakeya conjecture.
  • Results in quantum chromodynamics and Yang–Mills theory. The presenter emphasizes that the claimed Yang–Mills result addresses existence, not the mass gap, so the full Millennium Prize problem remains unsolved.
• Google SLEEP: A proposed training framework inspired by sleep. During a “wake” phase, an AI gathers experience; during a “sleep” phase, it consolidates information and practices on generated tasks. The video says early tests showed improved long-context recall, mathematical reasoning, and language learning, with less forgetting.
• Gemini 4 Carbon: Presented as a top-performing model across several benchmarks, including accuracy, hallucination resistance, virtual business management, and converting room photos into floor plans. It reportedly ranked less highly on software-development tasks but was cheaper than some leading alternatives. At the time of the video, access was limited to selected cyber defenders; the presenter says a full review could follow wider release.
• GPT-6.1 Terra: Described as a lower-cost middle tier between smaller, faster and larger, more capable models, with strengths in computer use and professional tasks. The presenter says it compares favorably with larger models on some benchmarks, while trailing them on others, particularly coding.
• Anthropic IPO filing: The presenter reviews figures attributed to Anthropic’s S-1 filing, including rapid revenue growth, a large accounting loss said to be mostly non-cash, and adjusted operating income. The video also notes major multiyear compute commitments, revenue concentration among a small number of large customers, and founders retaining voting control after an IPO. The valuation, timing, and other IPO details were described as unconfirmed.

Tools and research projects

• MotionBricks (NVIDIA): Generates smooth, real-time character or humanoid-robot motion from target poses or instructions. The video highlights combinations of actions, low latency, and released models for human characters and Unitree G1, with local-use instructions.
• IDOL: Converts one image into a 3D object using Gaussian splatting, filling in unseen areas. It offers adjustable detail and is described as small enough to run with 8 GB of VRAM; the presenter cites generation times of about 19 seconds on an RTX 4090 and 42 seconds on an RTX 3080.
• SceneBloom: Creates furnished 3D rooms from text prompts, aiming for sensible layouts and non-overlapping objects. Its code, LoRA adapters, and asset library were reported as available.
• Unlimited OCR 2: Baidu’s document-reading model processes multiple pages together to preserve context across long documents. The video says it performs well on text, tables, and complex documents and is relatively compact.
• UniAction: Generates video from an image, text, and structured action controls, such as force, camera movement, hand position, or robot pose. It is based on Wan 2.2; code, model, and training/evaluation scripts were reported as released.
• AutoSwarm: An open-source orchestration framework for coordinating multiple AI agents and running tasks in parallel across different agent systems. The presenter says it supports Windows, macOS, and Linux.
• Paradee 1.2: A very small text-to-speech model, described as an 8.5 MB compressed distillation of Kokoro. It runs on a CPU but offers only four preset voices and no voice cloning.
• Hy-SimulMT: A real-time Chinese–English translation model that can reproduce a speaker’s voice in the translated language. The video describes it as a 30-billion-parameter mixture-of-experts model, with smaller quantized versions available.
• NVIDIA SimReady Assets: A free collection of more than 2,600 simulation-ready 3D objects, with physical properties such as mass, friction, and collision information for robotics simulation and training.
• MatAnyone-E: Extracts video foregrounds while preserving fine details such as hair and translucent edges. It is presented as efficient enough for real-time 4K processing on an H200 and as needing roughly 3 GB of VRAM on consumer hardware; code and models were still pending.
• Open-H3-Regenerate-2K: Upscales generated video, including outputs from Minimax H3 and other models. The presenter notes that it can alter details, especially faces, despite adding resolution; local-use instructions were available.
• StreamVibe: A “proactive video agent” that observes a live task and decides when to offer guidance. Its examples include cooking and booking travel. Code and models were described as awaiting approval.
• MUNDO: Builds explorable, editable 3D environments from images. It represents objects and physical properties so users can simulate interactions and train robots; code and models were said to be planned for later release.
• Genie 4: Generates interactive worlds from prompts or images, with keyboard navigation and editable scenes. The video highlights newly added audio generation. Access was restricted to testers and selected subscribers or partners.
• PhysAg3D: Turns an object image into an interactive 3D model with movable parts and simulated physical behavior. The presenter says its code and model were released under the MIT license.
• Audio8 TTS: A small, multilingual speech model that can clone a voice from a short sample. The video describes it as a 0.6-billion-parameter model, around 2.6 GB, with Apache 2 licensing.
• SenseNova Code2Video: Generates animated explainers and motion graphics by writing and rendering code for each frame, making individual elements editable. Its model was described as large—about 55 GB.
• Energy-based image model: A proof of concept that differs from diffusion models: rather than removing noise, it learns an energy score for image quality and iteratively adjusts the image to lower that score. The presenter notes that its image quality is not yet competitive.

Open models highlighted

• OverSeer: A Meta world-model project designed to keep a shared environment consistent for multiple agents, using additional fixed-camera views to track the world. The video says a dataset was released, but not the model.
• Nanbeige 4.3 33B: Presented as a capable model for agent automation and professional tasks, though its full size requires substantial hardware.
• LFM3: A family of small models for text, vision, audio, multimodal interaction, and decision-making. The video emphasizes quantized versions and support for consumer hardware, including Apple chips.
• Apex 35B: A model aimed at research and scientific discovery, including writing and revising robot-control code. Quantized versions were said to be available, with the smallest around 9 GB.
• Groot: A family of mixture-of-experts models in several sizes. The presenter highlights the smaller version’s reported performance relative to larger models, alongside multiple quantization options.
• Clover 35B: A diffusion language model that generates and refines text in blocks rather than one token at a time. It is presented as fast and strong on agent and tool-use tests, but still large to run.
• Spark-X2.6-1.7B: A small, fast model intended for local devices. The presenter says it performs well against similarly sized models and supports NVIDIA GPUs, Ascend NPUs, and Apple Silicon.
• SenseNova-Loop 9B: Uses a recurrent/cyclic architecture that reuses layers to provide more computation without substantially increasing parameter count. It was presented as competitive on agent benchmarks.

Reviews, demonstrations, and practical guidance

The presenter demos many of the tools and compares selected models using benchmarks, while noting limitations such as access restrictions, model size, or fidelity issues. Several projects include code, weights, or instructions for downloading and running locally. In particular, the video points viewers to installation guidance for MotionBricks, IDOL, SceneBloom, Unlimited OCR 2, UniAction, AutoSwarm, Paradee 1.2, Hy-SimulMT, PhysAg3D, and various open models. The presenter also says they may review Gemini 4 Carbon more fully once it becomes broadly available.

A sponsored segment promotes Luma, highlighting virtual clothing fitting, Relight Studio, and AI agents that coordinate content-creation workflows on a shared canvas.

Main speakers and sources

• Main speaker: A single presenter/narrator from the AI Search channel; no other recurring speaker is identified.
• Sources discussed: OpenAI, Google, Anthropic, NVIDIA, Baidu, Tencent, Meta, Liquid AI, and the research teams behind the featured projects.
• Sponsor: Luma.